wip

2020-12-11 14:23:07 -08:00 · 2020-12-11 14:20:02 -08:00 · 2020-12-11 14:17:17 -08:00 · 2020-12-11 13:53:53 -08:00 · 2020-12-11 13:53:12 -08:00 · 2020-12-11 12:58:01 -08:00
334 changed files with 16456 additions and 3419 deletions
@@ -3,28 +3,6 @@ orbs:
    gcp-gke: circleci/gcp-gke@1.0.4
    go: circleci/go@1.3.0

-commands:
-  skip-job-on-doc-only-changes:
-    description: "Do not continue this job and exit with success for PRs with only doc changes"
-    steps:
-
-      - run:
-          name: docs-only changes skip check
-          command: |
-            # pipeline.git.base_revision is not always defined, so only proceed if all external vars are defined
-            if test -n "<< pipeline.git.base_revision >>" && test -n "<< pipeline.git.revision >>" && test -n "$(git diff --name-only << pipeline.git.base_revision >>...<< pipeline.git.revision >>)"
-            then
-                if git diff --name-only << pipeline.git.base_revision >>...<< pipeline.git.revision >> | egrep -qv '\.(md|rst)$'
-                then
-                    echo "Non-docs were modified in this PR, proceeding normally"
-                else
-                    echo "Only docs were modified in this PR, quitting this job"
-                    # disable skipping for now, as circleCI's base_revision is inconsistent leading to invalid ranges
-                    # circleci step halt
-                fi
-            else
-                echo "Can't perform skipping check w/o base_revision defined, continuing the job"
-            fi

 # TPU REFERENCES
 references:
@@ -85,278 +63,6 @@ references:


 jobs:
-    run_tests_torch_and_tf:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.6
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-torch_and_tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,torch,testing,sentencepiece]
-            - save_cache:
-                key: v0.4-{{ checksum "setup.py" }}
-                paths:
-                    - '~/.cache/pip'
-            - run: RUN_PT_TF_CROSS_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_tf ./tests/ -m is_pt_tf_cross_test --durations=0 | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_torch:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing,sentencepiece]
-            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_tf:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece]
-            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_tf ./tests/ | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_flax:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                keys:
-                    - v0.4-flax-{{ checksum "setup.py" }}
-                    - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: sudo pip install .[flax,sklearn,torch,testing,sentencepiece]
-            - save_cache:
-                  key: v0.4-flax-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_flax ./tests/ | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_pipelines_torch:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-torch-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing,sentencepiece]
-            - save_cache:
-                  key: v0.4-torch-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_torch -m is_pipeline_test ./tests/ | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_pipelines_tf:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-tf-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,testing,sentencepiece]
-            - save_cache:
-                  key: v0.4-tf-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_tf ./tests/ -m is_pipeline_test | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_custom_tokenizers:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        environment:
-            RUN_CUSTOM_TOKENIZERS: yes
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-custom_tokenizers-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[ja,testing,sentencepiece]
-            - run: python -m unidic download
-            - save_cache:
-                  key: v0.4-custom_tokenizers-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -s --make-reports=tests_custom_tokenizers ./tests/test_tokenization_bert_japanese.py | tee tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/tests_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_examples_torch:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.6
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - skip-job-on-doc-only-changes
-            - restore_cache:
-                  keys:
-                      - v0.4-torch_examples-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,sentencepiece,testing]
-            - run: pip install -r examples/requirements.txt
-            - save_cache:
-                  key: v0.4-torch_examples-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=examples_torch ./examples/ | tee examples_output.txt
-            - store_artifacts:
-                  path: ~/transformers/examples_output.txt
-            - store_artifacts:
-                  path: ~/transformers/reports
-
-    run_tests_git_lfs:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.7
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - run: sudo apt-get install git-lfs
-            - run: |
-                git config --global user.email "ci@dummy.com"
-                git config --global user.name "ci"
-            - run: pip install --upgrade pip
-            - run: pip install .[testing]
-            - run: RUN_GIT_LFS_TESTS=1 python -m pytest -sv ./tests/test_hf_api.py -k "HfLargefilesTest"
-
-    build_doc:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.6
-        steps:
-            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.4-build_doc-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install ."[all, docs]"
-            - save_cache:
-                  key: v0.4-build_doc-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: cd docs && make html SPHINXOPTS="-W"
-            - store_artifacts:
-                path: ./docs/_build
-
-    deploy_doc:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.6
-        steps:
-            - add_ssh_keys:
-                fingerprints:
-                    - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
-            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.4-deploy_doc-{{ checksum "setup.py" }}
-                      - v0.4-{{ checksum "setup.py" }}
-            - run: pip install ."[all,docs]"
-            - save_cache:
-                  key: v0.4-deploy_doc-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: ./.circleci/deploy.sh
-
    check_code_quality:
        working_directory: ~/transformers
        docker:
@@ -370,20 +76,22 @@ jobs:
                      - v0.4-code_quality-{{ checksum "setup.py" }}
                      - v0.4-{{ checksum "setup.py" }}
            - run: pip install --upgrade pip
-            - run: pip install isort
-            - run: pip install .[all,quality]
-            - save_cache:
-                  key: v0.4-code_quality-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: black --check examples tests src utils
-            - run: isort --check-only examples tests src utils
-            - run: flake8 examples tests src utils
-            - run: python utils/style_doc.py src/transformers docs/source --max_len 119 --check_only
-            - run: python utils/check_copies.py
-            - run: python utils/check_table.py
-            - run: python utils/check_dummies.py
-            - run: python utils/check_repo.py
+            - run:
+                command: |
+                    set +e
+                    echo "my experiment is here"
+                    # emulate failure
+                    false
+                    echo "it can safely fail"
+                    some non existing command
+                    echo "should still reach here"
+                    some non existing command again
+                    echo "should still reach here too"
+                    false
+            - run:
+                when: always
+                command: |
+                    echo "forcing success for this experiment"

    check_repository_consistency:
        working_directory: ~/transformers
@@ -395,37 +103,7 @@ jobs:
            - checkout
            - run: pip install requests
            - run: python ./utils/link_tester.py
-
-# TPU JOBS
-    run_examples_tpu:
-        docker:
-            - image: circleci/python:3.6
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - go/install
-            - *checkout_ml_testing
-            - gcp-gke/install
-            - gcp-gke/update-kubeconfig-with-credentials:
-                  cluster: $GKE_CLUSTER
-                  perform-login: true
-            - setup_remote_docker
-            - *build_push_docker
-            - *deploy_cluster
-
-    cleanup-gke-jobs:
-        docker:
-            - image: circleci/python:3.6
-        steps:
-            - gcp-gke/install
-            - gcp-gke/update-kubeconfig-with-credentials:
-                  cluster: $GKE_CLUSTER
-                  perform-login: true
-            - *delete_gke_jobs
-
+            
 workflow_filters: &workflow_filters
    filters:
        branches:
@@ -437,26 +115,5 @@ workflows:
        jobs:
            - check_code_quality
            - check_repository_consistency
-            - run_examples_torch
-            - run_tests_custom_tokenizers
-            - run_tests_torch_and_tf
-            - run_tests_torch
-            - run_tests_tf
-            - run_tests_flax
-            - run_tests_pipelines_torch
-            - run_tests_pipelines_tf
-            - run_tests_git_lfs
-            - build_doc
-            - deploy_doc: *workflow_filters
-    tpu_testing_jobs:
-        triggers:
-            - schedule:
-                # Set to run at the first minute of every hour.
-                cron: "0 8 * * *"
-                filters:
-                    branches:
-                        only:
-                            - master
-        jobs:
-            - cleanup-gke-jobs
-            - run_examples_tpu
+
+            
@@ -0,0 +1,65 @@
+name: Model templates runner
+
+on:
+  push:
+    paths:
+      - "src/**"
+      - "tests/**"
+      - ".github/**"
+      - "templates/**"
+
+jobs:
+  run_tests_templates:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v1
+
+      - name: Install Python
+        uses: actions/setup-python@v1
+        with:
+          python-version: 3.6
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: ~/.cache/pip
+          key: v1.2-tests_templates
+          restore-keys: |
+            v1.2-tests_templates-${{ hashFiles('setup.py') }}
+            v1.2-tests_templates
+
+      - name: Install dependencies
+        run: |
+          pip install --upgrade pip
+          pip install .[dev]
+      - name: Create model files
+        run: |
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/encoder-bert-tokenizer.json --path=templates/adding_a_new_model
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/pt-encoder-bert-tokenizer.json --path=templates/adding_a_new_model
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/standalone.json --path=templates/adding_a_new_model
+          transformers-cli add-new-model --testing --testing_file=templates/adding_a_new_model/tests/tf-encoder-bert-tokenizer.json --path=templates/adding_a_new_model
+          make style
+          python utils/check_table.py --fix_and_overwrite
+          python utils/check_dummies.py --fix_and_overwrite
+
+      - name: Run all non-slow tests
+        run: |
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_templates tests/*template*
+
+      - name: Run style changes
+        run: |
+          git fetch origin master:master
+          make fixup
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_templates_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_templates_test_reports
+          path: reports
@@ -159,3 +159,6 @@ tags

 # pre-commit
 .pre-commit*
+
+# .lock
+*.lock
@@ -213,6 +213,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[LXMERT](https://huggingface.co/transformers/model_doc/lxmert.html)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
 1. **[MarianMT](https://huggingface.co/transformers/model_doc/marian.html)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
 1. **[MBart](https://huggingface.co/transformers/model_doc/mbart.html)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
+1. **[MPNet](https://huggingface.co/transformers/model_doc/mpnet.html)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
 1. **[MT5](https://huggingface.co/transformers/model_doc/mt5.html)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
 1. **[Pegasus](https://huggingface.co/transformers/model_doc/pegasus.html)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[ProphetNet](https://huggingface.co/transformers/model_doc/prophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
@@ -1,10 +1,10 @@
 // These two things need to be updated at each release for the version selector.
 // Last stable version
 const stableVersion = "v4.0.0"
-// Dictionary doc folder to label
+// Dictionary doc folder to label. The last stable version should have an empty key.
 const versionMapping = {
    "master": "master",
-    "v4.0.0": "v4.0.0",
+    "": "v4.0.0 (stable)",
    "v3.5.1": "v3.5.0/v3.5.1",
    "v3.4.0": "v3.4.0",
    "v3.3.1": "v3.3.0/v3.3.1",
@@ -151,41 +151,44 @@ and conversion utilities for the following models:
 22. :doc:`MBart <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Denoising Pre-training for
    Neural Machine Translation <https://arxiv.org/abs/2001.08210>`__ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li,
    Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
-23. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
+23. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
+    Pre-training for Language Understanding <https://arxiv.org/abs/2004.09297>`__ by Kaitao Song, Xu Tan, Tao Qin,
+    Jianfeng Lu, Tie-Yan Liu.
+24. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
    text-to-text transformer <https://arxiv.org/abs/2010.11934>`__ by Linting Xue, Noah Constant, Adam Roberts, Mihir
    Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
-24. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
+25. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
    Gap-sentences for Abstractive Summarization <https://arxiv.org/abs/1912.08777>`__> by Jingqing Zhang, Yao Zhao,
    Mohammad Saleh and Peter J. Liu.
-25. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
+26. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
    Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan, Weizhen Qi,
    Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-26. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
+27. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
    Transformer <https://arxiv.org/abs/2001.04451>`__ by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
-27. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
+28. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
    Pretraining Approach <https://arxiv.org/abs/1907.11692>`__ by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar
    Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov. ultilingual BERT into `DistilmBERT
    <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German version of
    DistilBERT.
-28. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
+29. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
    about efficient neural networks? <https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi
    Krishna, and Kurt W. Keutzer.
-29. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
+30. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
    Unified Text-to-Text Transformer <https://arxiv.org/abs/1910.10683>`__ by Colin Raffel and Noam Shazeer and Adam
    Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
-30. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+31. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
    Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
    Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-31. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+32. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
    Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis Conneau.
-32. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+33. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
    Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
    Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-33. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+34. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
    Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
    Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
    Zettlemoyer and Veselin Stoyanov.
-34. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
+35. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
    Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
    Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.

@@ -240,6 +243,8 @@ TensorFlow and/or Flax.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         Longformer          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|            MPNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           Marian            |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         MobileBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
@@ -279,7 +284,6 @@ TensorFlow and/or Flax.
 |             mT5             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+

-
 .. toctree::
    :maxdepth: 2
    :caption: Get started
@@ -366,6 +370,7 @@ TensorFlow and/or Flax.
    model_doc/marian
    model_doc/mbart
    model_doc/mobilebert
+    model_doc/mpnet
    model_doc/mt5
    model_doc/gpt
    model_doc/gpt2
@@ -15,7 +15,8 @@ Utilities for Generation

 This page lists all the utility functions used by :meth:`~transformers.PretrainedModel.generate`,
 :meth:`~transformers.PretrainedModel.greedy_search`, :meth:`~transformers.PretrainedModel.sample`,
-:meth:`~transformers.PretrainedModel.beam_search`, and :meth:`~transformers.PretrainedModel.beam_sample`.
+:meth:`~transformers.PretrainedModel.beam_search`, :meth:`~transformers.PretrainedModel.beam_sample`, and
+:meth:`~transformers.PretrainedModel.group_beam_search`.

 Most of those are only useful if you are studying the code of the generate methods in the library.

@@ -31,6 +32,9 @@ generation.
 .. autoclass:: transformers.LogitsProcessorList
    :members: __call__

+.. autoclass:: transformers.LogitsWarper
+    :members: __call__
+
 .. autoclass:: transformers.MinLengthLogitsProcessor
    :members: __call__

@@ -52,6 +56,12 @@ generation.
 .. autoclass:: transformers.NoBadWordsLogitsProcessor
    :members: __call__

+.. autoclass:: transformers.PrefixConstrainedLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.HammingDiversityLogitsProcessor
+    :members: __call__
+
 BeamSearch
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -60,3 +70,10 @@ BeamSearch

 .. autoclass:: transformers.BeamSearchScorer
    :members: process, finalize
+
+Utilities
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autofunction:: transformers.top_k_top_p_filtering
+
+.. autofunction:: transformers.tf_top_k_top_p_filtering
@@ -22,6 +22,8 @@ Utilities

 .. autoclass:: transformers.EvalPrediction

+.. autoclass:: transformers.EvaluationStrategy
+
 .. autofunction:: transformers.set_seed

 .. autofunction:: transformers.torch_distributed_zero_first
@@ -32,8 +34,15 @@ Callbacks internals

 .. autoclass:: transformers.trainer_callback.CallbackHandler

+
 Distributed Evaluation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.trainer_pt_utils.DistributedTensorGatherer
    :members:
+
+
+Distributed Evaluation
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.HfArgumentParser
@@ -74,6 +74,10 @@ Learning Rate Schedules (Pytorch)
    :target: /imgs/warmup_linear_schedule.png
    :alt:

+
+.. autofunction:: transformers.get_polynomial_decay_schedule_with_warmup
+
+
 Warmup (TensorFlow)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

@@ -73,8 +73,9 @@ FillMaskPipeline
 NerPipeline
 =======================================================================================================================

-This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined below. Please refer to that
-pipeline for documentation and usage examples.
+.. autoclass:: transformers.NerPipeline
+
+See :class:`~transformers.TokenClassificationPipeline` for all details.

 QuestionAnsweringPipeline
 =======================================================================================================================
@@ -118,6 +119,13 @@ TokenClassificationPipeline
    :special-members: __call__
    :members:

+TranslationPipeline
+=======================================================================================================================
+
+.. autoclass:: transformers.TranslationPipeline
+    :special-members: __call__
+    :members:
+
 ZeroShotClassificationPipeline
 =======================================================================================================================

@@ -60,6 +60,13 @@ AlbertTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+AlbertTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AlbertTokenizerFast
+    :members:
+
+
 Albert specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -175,3 +175,10 @@ TFAutoModelForQuestionAnswering

 .. autoclass:: transformers.TFAutoModelForQuestionAnswering
    :members:
+
+
+FlaxAutoModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxAutoModel
+    :members:
@@ -98,6 +98,12 @@ BartTokenizer
    :members:


+BartTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartTokenizerFast
+    :members:
+

 BartModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -105,8 +111,6 @@ BartModel
 .. autoclass:: transformers.BartModel
    :members: forward

-.. autofunction:: transformers.models.bart.modeling_bart._prepare_bart_decoder_inputs
-

 BartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -51,3 +51,9 @@ BarthezTokenizer
 .. autoclass:: transformers.BarthezTokenizer
    :members:

+
+BarthezTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BarthezTokenizerFast
+    :members:
@@ -207,3 +207,10 @@ FlaxBertModel

 .. autoclass:: transformers.FlaxBertModel
    :members: __call__
+
+
+FlaxBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.FlaxBertForMaskedLM
+    :members: __call__
@@ -54,6 +54,13 @@ CamembertTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+CamembertTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.CamembertTokenizerFast
+    :members:
+
+
 CamembertModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -57,6 +57,13 @@ LayoutLMTokenizer
    :members:


+LayoutLMTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LayoutLMTokenizerFast
+    :members:
+
+
 LayoutLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -90,6 +90,13 @@ MBartTokenizer
    :members: build_inputs_with_special_tokens, prepare_seq2seq_batch


+MBartTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MBartTokenizerFast
+    :members:
+
+
 MBartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -0,0 +1,149 @@
+.. 
+    Copyright 2020 The HuggingFace Team. All rights reserved.
+
+    Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+    the License. You may obtain a copy of the License at
+
+        http://www.apache.org/licenses/LICENSE-2.0
+
+    Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+    an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+    specific language governing permissions and limitations under the License.
+
+MPNet
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The MPNet model was proposed in `MPNet: Masked and Permuted Pre-training for Language Understanding
+<https://arxiv.org/abs/2004.09297>`__ by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
+
+MPNet adopts a novel pre-training method, named masked and permuted language modeling, to inherit the advantages of
+masked language modeling and permuted language modeling for natural language understanding.
+
+The abstract from the paper is the following:
+
+*BERT adopts masked language modeling (MLM) for pre-training and is one of the most successful pre-training models.
+Since BERT neglects dependency among predicted tokens, XLNet introduces permuted language modeling (PLM) for
+pre-training to address this problem. However, XLNet does not leverage the full position information of a sentence and
+thus suffers from position discrepancy between pre-training and fine-tuning. In this paper, we propose MPNet, a novel
+pre-training method that inherits the advantages of BERT and XLNet and avoids their limitations. MPNet leverages the
+dependency among predicted tokens through permuted language modeling (vs. MLM in BERT), and takes auxiliary position
+information as input to make the model see a full sentence and thus reducing the position discrepancy (vs. PLM in
+XLNet). We pre-train MPNet on a large-scale dataset (over 160GB text corpora) and fine-tune on a variety of
+down-streaming tasks (GLUE, SQuAD, etc). Experimental results show that MPNet outperforms MLM and PLM by a large
+margin, and achieves better results on these tasks compared with previous state-of-the-art pre-trained methods (e.g.,
+BERT, XLNet, RoBERTa) under the same model setting.*
+
+Tips:
+
+- MPNet doesn't have :obj:`token_type_ids`, you don't need to indicate which token belongs to which segment. just
+  separate your segments with the separation token :obj:`tokenizer.sep_token` (or :obj:`[sep]`).
+
+The original code can be found `here <https://github.com/microsoft/MPNet>`__.
+
+MPNetConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetConfig
+    :members:
+
+
+MPNetTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+MPNetTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetTokenizerFast
+    :members:
+
+
+MPNetModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetModel
+    :members: forward
+
+
+MPNetForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetForMaskedLM
+    :members: forward
+
+
+MPNetForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetForSequenceClassification
+    :members: forward
+
+
+MPNetForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetForMultipleChoice
+    :members: forward
+
+
+MPNetForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetForTokenClassification
+    :members: forward
+
+
+MPNetForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MPNetForQuestionAnswering
+    :members: forward
+
+
+TFMPNetModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetModel
+    :members: call
+
+
+TFMPNetForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetForMaskedLM
+    :members: call
+
+
+TFMPNetForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetForSequenceClassification
+    :members: call
+
+
+TFMPNetForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetForMultipleChoice
+    :members: call
+
+
+TFMPNetForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetForTokenClassification
+    :members: call
+
+
+TFMPNetForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMPNetForQuestionAnswering
+    :members: call
@@ -37,6 +37,22 @@ MT5Config
    :members:


+MT5Tokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MT5Tokenizer
+
+See :class:`~transformers.T5Tokenizer` for all details.
+
+
+MT5TokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MT5TokenizerFast
+
+See :class:`~transformers.T5TokenizerFast` for all details.
+
+
 MT5Model
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -112,6 +112,13 @@ warning: ``add_tokens`` does not work at the moment.
    :members: __call__, prepare_seq2seq_batch


+PegasusTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.PegasusTokenizerFast
+    :members:
+
+
 PegasusForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -163,6 +163,13 @@ ReformerTokenizer
    :members: save_vocabulary


+ReformerTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ReformerTokenizerFast
+    :members:
+
+
 ReformerModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -107,6 +107,13 @@ T5Tokenizer
        create_token_type_ids_from_sequences, prepare_seq2seq_batch, save_vocabulary


+T5TokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.T5TokenizerFast
+    :members:
+
+
 T5Model
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -105,3 +105,11 @@ TFTransfoXLLMHeadModel

 .. autoclass:: transformers.TFTransfoXLLMHeadModel
    :members: call
+
+
+Internal Layers
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.AdaptiveEmbedding
+
+.. autoclass:: transformers.TFAdaptiveEmbedding
@@ -62,6 +62,13 @@ XLMRobertaTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+XLMRobertaTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLMRobertaTokenizerFast
+    :members:
+
+
 XLMRobertaModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -62,6 +62,13 @@ XLNetTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


+XLNetTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.XLNetTokenizerFast
+    :members:
+
+
 XLNet specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -16,59 +16,58 @@ limitations under the License.

 # Examples

-Version 2.9 of 🤗 Transformers introduced a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.
-
-Here is the list of all our examples:
- **grouped by task** (all official examples work for multiple models)
- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
-  just lack some features),
- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library.
- links to **Colab notebooks** to walk through the scripts and run them easily,
- links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.
-
+This folder contains actively maintained examples of use of 🤗 Transformers organized along NLP tasks. If you are looking for an example that used to
+be in this folder, it may have moved to our [research projects](https://github.com/huggingface/transformers/tree/master/examples/research_projects) subfolder (which contains frozen snapshots of research projects).

 ## Important note

 **Important**

-To make sure you can successfully run the latest versions of the example scripts, you have to **install the library from source** and install some example-specific requirements.
-Execute the following steps in a new virtual environment:
-
+To make sure you can successfully run the latest versions of the example scripts, you have to **install the library from source** and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
 ```bash
 git clone https://github.com/huggingface/transformers
 cd transformers
 pip install .
-pip install -r ./examples/requirements.txt
+```
+Then cd in the example folder of your choice and run
+```bash
+pip install -r requirements.txt
 ```

-Alternatively, you can run the version of the examples as they were for your current version of Transformers via (for instance with v3.4.0):
+Alternatively, you can run the version of the examples as they were for your current version of Transformers via (for instance with v3.5.1):
 ```bash
-git checkout tags/v3.4.0
+git checkout tags/v3.5.1
 ```

 ## The Big Table of Tasks

+Here is the list of all our examples:
+- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
+  just lack some features),
+- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library.
+- links to **Colab notebooks** to walk through the scripts and run them easily,
+<!--
+Coming soon!
+- links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.
+-->
+
 | Task | Example datasets | Trainer support | TFTrainer support | 🤗 Datasets | Colab
 |---|---|:---:|:---:|:---:|:---:|
 | [**`language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/language-modeling)       | Raw text        | ✅ | -  | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
-| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb)
-| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
 | [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
-| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | - | -
-| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
-| [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)                 | All             | - | -  | - | -
+| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | ✅ | -
 | [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                     | CNN/Daily Mail  | ✅  | - | - | -
+| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb)
+| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | - | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
+| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
 | [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | - | - | -
-| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | - | - | -
-| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | - | - | -


-<br>
-
+<!--
 ## One-click Deploy to Cloud (wip)

 **Coming soon!**
+-->

 ## Running on TPUs

@@ -0,0 +1,20 @@
+tensorboard
+scikit-learn
+seqeval
+psutil
+sacrebleu
+rouge-score
+tensorflow_datasets
+matplotlib
+git-python==1.0.3
+faiss-cpu
+streamlit
+elasticsearch
+nltk
+pandas
+datasets >= 1.1.3
+fire
+pytest
+conllu
+sentencepiece != 0.1.92
+protobuf
@@ -1,3 +1,19 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
 # 🤗 Benchmark results

 Here, you can find a list of the different benchmark results created by the community.
@@ -1,3 +1,17 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import csv
 from collections import defaultdict
 from dataclasses import dataclass, field
@@ -1,3 +1,17 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 # tests directory-specific settings - this file is run automatically
 # by pytest before any tests are run

@@ -1,5 +0,0 @@
-# Community contributed examples
-
-This folder contains examples which are not actively maintained (mostly contributed by the community).
-
-Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.
@@ -1,3 +1,19 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
 ## Language model training

 Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2,
@@ -0,0 +1,3 @@
+datasets >= 1.1.3
+sentencepiece != 0.1.92
+protobuf
@@ -0,0 +1,636 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=masked-lm
+"""
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import jax
+import jax.numpy as jnp
+from flax import jax_utils
+from flax.optim import Adam
+from flax.training import common_utils
+from flax.training.common_utils import get_metrics
+from jax.nn import log_softmax
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoTokenizer,
+    FlaxBertForMaskedLM,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+
+
+# Cache the result
+has_tensorboard = is_tensorboard_available()
+if has_tensorboard:
+    try:
+        from flax.metrics.tensorboard import SummaryWriter
+    except ImportError as ie:
+        has_tensorboard = False
+        print(f"Unable to display metrics through TensorBoard because some package are not installed: {ie}")
+
+else:
+    print(
+        "Unable to display metrics through TensorBoard because the package is not installed: "
+        "Please run pip install tensorboard to enable."
+    )
+
+
+MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+# Adapted from transformers/data/data_collator.py
+# Letting here for now, let's discuss where it should live
+@dataclass
+class FlaxDataCollatorForLanguageModeling:
+    """
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to use masked language modeling. If set to :obj:`False`, the labels are the same as the
+            inputs with the padding tokens ignored (by setting them to -100). Otherwise, the labels are -100 for
+            non-masked tokens and the value to predict for the masked token.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input, when :obj:`mlm` is set to :obj:`True`.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    mlm: bool = True
+    mlm_probability: float = 0.15
+
+    def __post_init__(self):
+        if self.mlm and self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
+    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
+        # Handle dict or lists with proper padding and conversion to tensor.
+        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+        if self.mlm:
+            batch["input_ids"], batch["labels"] = self.mask_tokens(
+                batch["input_ids"], special_tokens_mask=special_tokens_mask
+            )
+        else:
+            labels = batch["input_ids"].copy()
+            if self.tokenizer.pad_token_id is not None:
+                labels[labels == self.tokenizer.pad_token_id] = -100
+            batch["labels"] = labels
+        return batch
+
+    def mask_tokens(
+        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
+    ) -> Tuple[jnp.ndarray, jnp.ndarray]:
+        """
+        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
+        """
+        labels = inputs.copy()
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
+        probability_matrix = np.full(labels.shape, self.mlm_probability)
+        special_tokens_mask = special_tokens_mask.astype("bool")
+
+        probability_matrix[special_tokens_mask] = 0.0
+        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
+        labels[~masked_indices] = -100  # We only compute loss on masked tokens
+
+        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
+        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
+
+        # 10% of the time, we replace masked input tokens with random word
+        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
+        indices_random &= masked_indices & ~indices_replaced
+
+        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
+        inputs[indices_random] = random_words[indices_random]
+
+        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+        return inputs, labels
+
+
+def create_learning_rate_scheduler(
+    factors="constant * linear_warmup * rsqrt_decay",
+    base_learning_rate=0.5,
+    warmup_steps=1000,
+    decay_factor=0.5,
+    steps_per_decay=20000,
+    steps_per_cycle=100000,
+):
+    """Creates learning rate schedule.
+    Interprets factors in the factors string which can consist of:
+    * constant: interpreted as the constant value,
+    * linear_warmup: interpreted as linear warmup until warmup_steps,
+    * rsqrt_decay: divide by square root of max(step, warmup_steps)
+    * rsqrt_normalized_decay: divide by square root of max(step/warmup_steps, 1)
+    * decay_every: Every k steps decay the learning rate by decay_factor.
+    * cosine_decay: Cyclic cosine decay, uses steps_per_cycle parameter.
+    Args:
+      factors: string, factors separated by "*" that defines the schedule.
+      base_learning_rate: float, the starting constant for the lr schedule.
+      warmup_steps: int, how many steps to warm up for in the warmup schedule.
+      decay_factor: float, the amount to decay the learning rate by.
+      steps_per_decay: int, how often to decay the learning rate.
+      steps_per_cycle: int, steps per cycle when using cosine decay.
+    Returns:
+      a function learning_rate(step): float -> {"learning_rate": float}, the
+      step-dependent lr.
+    """
+    factors = [n.strip() for n in factors.split("*")]
+
+    def step_fn(step):
+        """Step to learning rate function."""
+        ret = 1.0
+        for name in factors:
+            if name == "constant":
+                ret *= base_learning_rate
+            elif name == "linear_warmup":
+                ret *= jnp.minimum(1.0, step / warmup_steps)
+            elif name == "rsqrt_decay":
+                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
+            elif name == "rsqrt_normalized_decay":
+                ret *= jnp.sqrt(warmup_steps)
+                ret /= jnp.sqrt(jnp.maximum(step, warmup_steps))
+            elif name == "decay_every":
+                ret *= decay_factor ** (step // steps_per_decay)
+            elif name == "cosine_decay":
+                progress = jnp.maximum(0.0, (step - warmup_steps) / float(steps_per_cycle))
+                ret *= jnp.maximum(0.0, 0.5 * (1.0 + jnp.cos(jnp.pi * (progress % 1.0))))
+            else:
+                raise ValueError("Unknown factor %s." % name)
+        return jnp.asarray(ret, dtype=jnp.float32)
+
+    return step_fn
+
+
+def compute_metrics(logits, labels, weights, label_smoothing=0.0):
+    """Compute summary metrics."""
+    loss, normalizer = cross_entropy(logits, labels, weights, label_smoothing)
+    acc, _ = accuracy(logits, labels, weights)
+    metrics = {"loss": loss, "accuracy": acc, "normalizer": normalizer}
+    metrics = jax.lax.psum(metrics, axis_name="batch")
+    return metrics
+
+
+def accuracy(logits, targets, weights=None):
+    """Compute weighted accuracy for log probs and targets.
+    Args:
+     logits: [batch, length, num_classes] float array.
+     targets: categorical targets [batch, length] int array.
+     weights: None or array of shape [batch, length]
+    Returns:
+      Tuple of scalar loss and batch normalizing factor.
+    """
+    if logits.ndim != targets.ndim + 1:
+        raise ValueError(
+            "Incorrect shapes. Got shape %s logits and %s targets" % (str(logits.shape), str(targets.shape))
+        )
+
+    loss = jnp.equal(jnp.argmax(logits, axis=-1), targets)
+    loss *= weights
+
+    return loss.sum(), weights.sum()
+
+
+def cross_entropy(logits, targets, weights=None, label_smoothing=0.0):
+    """Compute cross entropy and entropy for log probs and targets.
+    Args:
+     logits: [batch, length, num_classes] float array.
+     targets: categorical targets [batch, length] int array.
+     weights: None or array of shape [batch, length]
+     label_smoothing: label smoothing constant, used to determine the on and off values.
+    Returns:
+      Tuple of scalar loss and batch normalizing factor.
+    """
+    if logits.ndim != targets.ndim + 1:
+        raise ValueError(
+            "Incorrect shapes. Got shape %s logits and %s targets" % (str(logits.shape), str(targets.shape))
+        )
+
+    vocab_size = logits.shape[-1]
+    confidence = 1.0 - label_smoothing
+    low_confidence = (1.0 - confidence) / (vocab_size - 1)
+    normalizing_constant = -(
+        confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
+    )
+    soft_targets = common_utils.onehot(targets, vocab_size, on_value=confidence, off_value=low_confidence)
+
+    loss = -jnp.sum(soft_targets * log_softmax(logits), axis=-1)
+    loss = loss - normalizing_constant
+
+    if weights is not None:
+        loss = loss * weights
+        normalizing_factor = weights.sum()
+    else:
+        normalizing_factor = np.prod(targets.shape)
+
+    return loss.sum(), normalizing_factor
+
+
+def training_step(optimizer, batch, dropout_rng):
+    dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+    def loss_fn(params):
+        targets = batch.pop("labels")
+
+        # Hide away tokens which doesn't participate in the optimization
+        token_mask = jnp.where(targets > 0, 1.0, 0.0)
+
+        pooled, logits = model(**batch, params=params, dropout_rng=dropout_rng, train=True)
+        loss, weight_sum = cross_entropy(logits, targets, token_mask)
+        return loss / weight_sum
+
+    step = optimizer.state.step
+    lr = lr_scheduler_fn(step)
+    grad_fn = jax.value_and_grad(loss_fn)
+    loss, grad = grad_fn(optimizer.target)
+    grad = jax.lax.pmean(grad, "batch")
+    optimizer = optimizer.apply_gradient(grad, learning_rate=lr)
+
+    return loss, optimizer, new_dropout_rng
+
+
+def eval_step(params, batch):
+    """
+    Calculate evaluation metrics on a batch.
+    """
+    targets = batch.pop("labels")
+
+    # Hide away tokens which doesn't participate in the optimization
+    token_mask = jnp.where(targets > 0, 1.0, 0.0)
+    _, logits = model(**batch, params=params, train=False)
+
+    return compute_metrics(logits, targets, token_mask)
+
+
+def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    nb_samples = len(samples_idx)
+    samples_to_remove = nb_samples % batch_size
+
+    if samples_to_remove != 0:
+        samples_idx = samples_idx[:-samples_to_remove]
+    sections_split = nb_samples // batch_size
+    batch_idx = jnp.split(samples_idx, sections_split)
+    return batch_idx
+
+
+if __name__ == "__main__":
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        level="NOTSET",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        if extension == "txt":
+            extension = "text"
+        datasets = load_dataset(extension, data_files=data_files)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    padding = "max_length" if data_args.pad_to_max_length else False
+
+    def tokenize_function(examples):
+        # Remove empty lines
+        examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
+        return tokenizer(
+            examples["text"],
+            return_special_tokens_mask=True,
+            padding=padding,
+            truncation=True,
+            max_length=data_args.max_seq_length,
+        )
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=[text_column_name],
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Enable tensorboard only on the master node
+    if has_tensorboard and jax.host_id() == 0:
+        summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir).joinpath("logs").as_posix())
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our training
+    rng = jax.random.PRNGKey(training_args.seed)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    model = FlaxBertForMaskedLM.from_pretrained("bert-base-cased", dtype=jnp.float32, dropout_rate=0.1)
+    model.init(jax.random.PRNGKey(training_args.seed), (training_args.train_batch_size, model.config.max_length))
+
+    # Setup optimizer
+    optimizer = Adam(
+        learning_rate=training_args.learning_rate,
+        weight_decay=training_args.weight_decay,
+        beta1=training_args.adam_beta1,
+        beta2=training_args.adam_beta2,
+    ).create(model.params)
+
+    # Create learning rate scheduler
+    lr_scheduler_fn = create_learning_rate_scheduler(
+        base_learning_rate=training_args.learning_rate, warmup_steps=training_args.warmup_steps
+    )
+
+    # Create parallel version of the training and evaluation steps
+    p_training_step = jax.pmap(training_step, "batch", donate_argnums=(0,))
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the optimizer on each device
+    optimizer = jax_utils.replicate(optimizer)
+
+    # Store some constant
+    nb_epochs = int(training_args.num_train_epochs)
+    batch_size = int(training_args.train_batch_size)
+    eval_batch_size = int(training_args.eval_batch_size)
+
+    epochs = tqdm(range(nb_epochs), desc=f"Epoch ... (1/{nb_epochs})", position=0)
+    for epoch in epochs:
+
+        # ======================== Training ================================
+        # Create sampling rng
+        rng, training_rng, eval_rng = jax.random.split(rng, 3)
+
+        # Generate an epoch by shuffling sampling indices from the train dataset
+        nb_training_samples = len(tokenized_datasets["train"])
+        training_samples_idx = jax.random.permutation(training_rng, jnp.arange(nb_training_samples))
+        training_batch_idx = generate_batch_splits(training_samples_idx, batch_size)
+
+        # Gather the indexes for creating the batch and do a training step
+        for batch_idx in tqdm(training_batch_idx, desc="Training...", position=1):
+            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = common_utils.shard(model_inputs.data)
+            loss, optimizer, dropout_rngs = p_training_step(optimizer, model_inputs, dropout_rngs)
+
+        epochs.write(f"Loss: {loss}")
+
+        # ======================== Evaluating ==============================
+        nb_eval_samples = len(tokenized_datasets["test"])
+        eval_samples_idx = jnp.arange(nb_eval_samples)
+        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+        eval_metrics = []
+        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+            samples = [tokenized_datasets["test"][int(idx)] for idx in batch_idx]
+            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+
+            # Model forward
+            model_inputs = common_utils.shard(model_inputs.data)
+            metrics = p_eval_step(optimizer.target, model_inputs)
+            eval_metrics.append(metrics)
+
+        eval_metrics_np = get_metrics(eval_metrics)
+        eval_metrics_np = jax.tree_map(jnp.sum, eval_metrics_np)
+        eval_normalizer = eval_metrics_np.pop("normalizer")
+        eval_summary = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics_np)
+
+        # Update progress bar
+        epochs.desc = (
+            f"Epoch... ({epoch + 1}/{nb_epochs} | Loss: {eval_summary['loss']}, Acc: {eval_summary['accuracy']})"
+        )
+
+        # Save metrics
+        if has_tensorboard and jax.host_id() == 0:
+            for name, value in eval_summary.items():
+                summary_writer.scalar(name, value, epoch)
@@ -281,7 +281,7 @@ def main():
    # Add the chinese references if provided
    if data_args.train_ref_file is not None:
        tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
-    if data_args.valid_ref_file is not None:
+    if data_args.validation_ref_file is not None:
        tokenized_datasets["validation"] = add_chinese_references(
            tokenized_datasets["validation"], data_args.validation_ref_file
        )
@@ -0,0 +1,21 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# Legacy examples
+
+This folder contains examples which are not actively maintained (mostly contributed by the community).
+
+Using these examples together with a recent version of the library usually requires to make small (sometimes big) adaptations to get the scripts working.
@@ -21,7 +21,7 @@ mkdir -p $OUTPUT_DIR
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_glue.py --gpus 1 --data_dir $DATA_DIR \
+python3 run_glue.py --gpus 1 --data_dir $DATA_DIR \
 --task $TASK \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -31,7 +31,7 @@ mkdir -p $OUTPUT_DIR
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_ner.py --data_dir ./ \
+python3 run_ner.py --data_dir ./ \
 --labels ./labels.txt \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -26,7 +26,7 @@ export SEED=1
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"

-python3 run_pl_ner.py --data_dir ./ \
+python3 run_ner.py --data_dir ./ \
 --task_type POS \
 --model_name_or_path $BERT_MODEL \
 --output_dir $OUTPUT_DIR \
@@ -0,0 +1,229 @@
+## Token classification
+
+Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/token-classification/run_ner.py).
+
+The following examples are covered in this section:
+
+* NER on the GermEval 2014 (German NER) dataset
+* Emerging and Rare Entities task: WNUT’17 (English NER) dataset
+
+Details and results for the fine-tuning provided by @stefan-it.
+
+### GermEval 2014 (German NER) dataset
+
+#### Data (Download and pre-processing steps)
+
+Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
+
+Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
+
+```bash
+curl -L 'https://drive.google.com/uc?export=download&id=1Jjhbal535VVz2ap4v4r_rN1UEHTdLK5P' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1ZfRcQThdtAR5PPRjIDtrVP7BtXSCUBbm' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://drive.google.com/uc?export=download&id=1u9mb7kNJHWQCWyweMDRMuTFoOHOfeBTH' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+```
+
+The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`.
+One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s.
+The `preprocess.py` script located in the `scripts` folder a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
+
+Let's define some variables that we need for further pre-processing steps and training the model:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+```
+
+Run the pre-processing script on training, dev and test datasets:
+
+```bash
+python3 scripts/preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 scripts/preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 scripts/preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+```
+
+The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
+
+```bash
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+```
+
+#### Prepare the run
+
+Additional environment variables must be set:
+
+```bash
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+```
+
+#### Run the Pytorch version
+
+To start training, just run:
+
+```bash
+python3 run_ner.py --data_dir ./ \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_device_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+```
+
+If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
+
+#### JSON-based configuration file
+
+Instead of passing all parameters via commandline arguments, the `run_ner.py` script also supports reading parameters from a json-based configuration file:
+
+```json
+{
+    "data_dir": ".",
+    "labels": "./labels.txt",
+    "model_name_or_path": "bert-base-multilingual-cased",
+    "output_dir": "germeval-model",
+    "max_seq_length": 128,
+    "num_train_epochs": 3,
+    "per_device_train_batch_size": 32,
+    "save_steps": 750,
+    "seed": 1,
+    "do_train": true,
+    "do_eval": true,
+    "do_predict": true
+}
+```
+
+It must be saved with a `.json` extension and can be used by running `python3 run_ner.py config.json`.
+
+#### Evaluation
+
+Evaluation on development dataset outputs the following for our example:
+
+```bash
+10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
+10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
+10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
+10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
+10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
+10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
+10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
+```
+
+### Emerging and Rare Entities task: WNUT’17 (English NER) dataset
+
+Description of the WNUT’17 task from the [shared task website](http://noisy-text.github.io/2017/index.html):
+
+> The WNUT’17 shared task focuses on identifying unusual, previously-unseen entities in the context of emerging discussions.
+> Named entities form the basis of many modern approaches to other tasks (like event clustering and summarization), but recall on
+> them is a real problem in noisy text - even among annotators. This drop tends to be due to novel entities and surface forms.
+
+Six labels are available in the dataset. An overview can be found on this [page](http://noisy-text.github.io/2017/files/).
+
+#### Data (Download and pre-processing steps)
+
+The dataset can be downloaded from the [official GitHub](https://github.com/leondz/emerging_entities_17) repository.
+
+The following commands show how to prepare the dataset for fine-tuning:
+
+```bash
+mkdir -p data_wnut_17
+
+curl -L 'https://github.com/leondz/emerging_entities_17/raw/master/wnut17train.conll'  | tr '\t' ' ' > data_wnut_17/train.txt.tmp
+curl -L 'https://github.com/leondz/emerging_entities_17/raw/master/emerging.dev.conll' | tr '\t' ' ' > data_wnut_17/dev.txt.tmp
+curl -L 'https://raw.githubusercontent.com/leondz/emerging_entities_17/master/emerging.test.annotated' | tr '\t' ' ' > data_wnut_17/test.txt.tmp
+```
+
+Let's define some variables that we need for further pre-processing steps:
+
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-large-cased
+```
+
+Here we use the English BERT large model for fine-tuning.
+The `preprocess.py` scripts splits longer sentences into smaller ones (once the max. subtoken length is reached):
+
+```bash
+python3 scripts/preprocess.py data_wnut_17/train.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/train.txt
+python3 scripts/preprocess.py data_wnut_17/dev.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/dev.txt
+python3 scripts/preprocess.py data_wnut_17/test.txt.tmp $BERT_MODEL $MAX_LENGTH > data_wnut_17/test.txt
+```
+
+In the last pre-processing step, the `labels.txt` file needs to be generated. This file contains all available labels:
+
+```bash
+cat data_wnut_17/train.txt data_wnut_17/dev.txt data_wnut_17/test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > data_wnut_17/labels.txt
+```
+
+#### Run the Pytorch version
+
+Fine-tuning with the PyTorch version can be started using the `run_ner.py` script. In this example we use a JSON-based configuration file.
+
+This configuration file looks like:
+
+```json
+{
+    "data_dir": "./data_wnut_17",
+    "labels": "./data_wnut_17/labels.txt",
+    "model_name_or_path": "bert-large-cased",
+    "output_dir": "wnut-17-model-1",
+    "max_seq_length": 128,
+    "num_train_epochs": 3,
+    "per_device_train_batch_size": 32,
+    "save_steps": 425,
+    "seed": 1,
+    "do_train": true,
+    "do_eval": true,
+    "do_predict": true,
+    "fp16": false
+}
+```
+
+If your GPU supports half-precision training, please set `fp16` to `true`.
+
+Save this JSON-based configuration under `wnut_17.json`. The fine-tuning can be started with `python3 run_ner_old.py wnut_17.json`.
+
+#### Evaluation
+
+Evaluation on development dataset outputs the following:
+
+```bash
+05/29/2020 23:33:44 - INFO - __main__ -   ***** Eval results *****
+05/29/2020 23:33:44 - INFO - __main__ -     eval_loss = 0.26505235286212275
+05/29/2020 23:33:44 - INFO - __main__ -     eval_precision = 0.7008264462809918
+05/29/2020 23:33:44 - INFO - __main__ -     eval_recall = 0.507177033492823
+05/29/2020 23:33:44 - INFO - __main__ -     eval_f1 = 0.5884802220680084
+05/29/2020 23:33:44 - INFO - __main__ -     epoch = 3.0
+```
+
+On the test dataset the following results could be achieved:
+
+```bash
+05/29/2020 23:33:44 - INFO - transformers.trainer -   ***** Running Prediction *****
+05/29/2020 23:34:02 - INFO - __main__ -     eval_loss = 0.30948806500973547
+05/29/2020 23:34:02 - INFO - __main__ -     eval_precision = 0.5840108401084011
+05/29/2020 23:34:02 - INFO - __main__ -     eval_recall = 0.3994439295644115
+05/29/2020 23:34:02 - INFO - __main__ -     eval_f1 = 0.47440836543753434
+```
+
+WNUT’17 is a very difficult task. Current state-of-the-art results on this dataset can be found [here](http://nlpprogress.com/english/named_entity_recognition.html).
@@ -20,7 +20,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type NER \
 --data_dir . \
 --labels ./labels.txt \
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type Chunk \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -21,7 +21,7 @@ export NUM_EPOCHS=3
 export SAVE_STEPS=750
 export SEED=1

-python3 run_ner_old.py \
+python3 run_ner.py \
 --task_type POS \
 --data_dir . \
 --model_name_or_path $BERT_MODEL \
@@ -1,3 +1,19 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
 ## Multiple Choice

 Based on the script [`run_multiple_choice.py`]().
@@ -0,0 +1,2 @@
+sentencepiece != 0.1.92
+protobuf
@@ -1,39 +1,44 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.

+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->

 ## SQuAD

-Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/question-answering/run_squad.py).
+Based on the script [`run_qa.py`](https://github.com/huggingface/transformers/blob/master/examples/question-answering/run_qa.py).
+
+**Note:** This script only works with models that have a fast tokenizer (backed by the 🤗 Tokenizers library) as it
+uses special features of those tokenizers. You can check if your favorite model has a fast tokenizer in
+[this table](https://huggingface.co/transformers/index.html#bigtable), if it doesn't you can still use the old version
+of the script.
+
+The old version of this script can be found [here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/question-answering/run_squad.py).

 #### Fine-tuning BERT on SQuAD1.0

 This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
-on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a
-$SQUAD_DIR directory.
-
-* [train-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json)
-* [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
-* [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
-
-And for SQuAD2.0, you need to download:
-
- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
+on a single tesla V100 16GB.

 ```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-  --model_type bert \
+python run_qa.py \
  --model_name_or_path bert-base-uncased \
+  --dataset_name squad \
  --do_train \
  --do_eval \
-  --do_lower_case \
-  --train_file $SQUAD_DIR/train-v1.1.json \
-  --predict_file $SQUAD_DIR/dev-v1.1.json \
-  --per_gpu_train_batch_size 12 \
+  --per_device_train_batch_size 12 \
  --learning_rate 3e-5 \
-  --num_train_epochs 2.0 \
+  --num_train_epochs 2 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir /tmp/debug_squad/
@@ -53,20 +58,17 @@ Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word

 ```bash
 python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
-    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --dataset_name squad \
    --do_train \
    --do_eval \
-    --do_lower_case \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=3   \
-    --per_gpu_train_batch_size=3   \
+    --per_device_eval_batch_size=3   \
+    --per_device_train_batch_size=3   \
 ```

 Training with the previously defined hyper-parameters yields the following results:
@@ -79,29 +81,25 @@ exact_match = 86.91
 This fine-tuned model is available as a checkpoint under the reference
 [`bert-large-uncased-whole-word-masking-finetuned-squad`](https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad).

-#### Fine-tuning XLNet on SQuAD
+#### Fine-tuning XLNet with beam search on SQuAD

-This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
+This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset.

 ##### Command for SQuAD1.0:

 ```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-    --model_type xlnet \
+python run_qa_beam_search.py \
    --model_name_or_path xlnet-large-cased \
+    --dataset_name squad \
    --do_train \
    --do_eval \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=4  \
-    --per_gpu_train_batch_size=4   \
+    --per_device_eval_batch_size=4  \
+    --per_device_train_batch_size=4   \
    --save_steps 5000
 ```

@@ -110,21 +108,19 @@ python run_squad.py \
 ```bash
 export SQUAD_DIR=/path/to/SQUAD

-python run_squad.py \
-    --model_type xlnet \
+python run_qa_beam_search.py \
    --model_name_or_path xlnet-large-cased \
+    --dataset_name squad_v2 \
    --do_train \
    --do_eval \
    --version_2_with_negative \
-    --train_file $SQUAD_DIR/train-v2.0.json \
-    --predict_file $SQUAD_DIR/dev-v2.0.json \
    --learning_rate 3e-5 \
    --num_train_epochs 4 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=2  \
-    --per_gpu_train_batch_size=2   \
+    --per_device_eval_batch_size=2  \
+    --per_device_train_batch_size=2   \
    --save_steps 5000
 ```

@@ -162,7 +158,7 @@ Larger batch size may improve the performance while costing more memory.
 #### Fine-tuning BERT on SQuAD1.0 with relative position embeddings

 The following examples show how to fine-tune BERT models with different relative position embeddings. The BERT model 
-`bert-base-uncased` was pre-trained with default absolute position embeddings. We provide the following pre-trained 
+`bert-base-uncased` was pretrained with default absolute position embeddings. We provide the following pretrained 
 models which were pre-trained on the same training data (BooksCorpus and English Wikipedia) as in the BERT model 
 training, but with different relative position embeddings. 

@@ -178,24 +174,19 @@ in Huang et al. [Improve Transformer Models with Better Relative Position Embedd
 ##### Base models fine-tuning

 ```bash
-export SQUAD_DIR=/path/to/SQUAD
-output_dir=relative_squad
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
-    --model_type bert \
    --model_name_or_path zhiheng-huang/bert-base-uncased-embedding-relative-key-query \
+    --dataset_name squad \
    --do_train \
    --do_eval \
-    --do_lower_case \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 512 \
    --doc_stride 128 \
-    --output_dir ${output_dir} \
-    --per_gpu_eval_batch_size=60 \
-    --per_gpu_train_batch_size=6
+    --output_dir relative_squad \
+    --per_device_eval_batch_size=60 \
+    --per_device_train_batch_size=6
 ```
 Training with the above command leads to the following results. It boosts the BERT default from f1 score of 88.52 to 90.54.

@@ -211,22 +202,17 @@ gpu training leads to the f1 score of 90.71.
 ##### Large models fine-tuning

 ```bash
-export SQUAD_DIR=/path/to/SQUAD
-output_dir=relative_squad
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
-    --model_type bert \
    --model_name_or_path zhiheng-huang/bert-large-uncased-whole-word-masking-embedding-relative-key-query \
+    --dataset_name squad \
    --do_train \
    --do_eval \
-    --do_lower_case \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 512 \
    --doc_stride 128 \
-    --output_dir ${output_dir} \
+    --output_dir relative_squad \
    --per_gpu_eval_batch_size=6 \
    --per_gpu_train_batch_size=2 \
    --gradient_accumulation_steps 3
@@ -251,5 +237,4 @@ python run_tf_squad.py \
    --doc_stride 128    
 ```

-
 For the moment evaluation is not available in the Tensorflow Trainer only the training.
@@ -0,0 +1 @@
+datasets >= 1.1.3
@@ -0,0 +1,469 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for question answering.
+"""
+# You can also adapt this script on your own question answering task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import load_dataset, load_metric
+
+import transformers
+from trainer_qa import QuestionAnsweringTrainer
+from transformers import (
+    AutoConfig,
+    AutoModelForQuestionAnswering,
+    AutoTokenizer,
+    DataCollatorWithPadding,
+    EvalPrediction,
+    HfArgumentParser,
+    PreTrainedTokenizerFast,
+    TrainingArguments,
+    default_data_collator,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+from utils_qa import postprocess_qa_predictions
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to directory to store the pretrained models downloaded from huggingface.co"},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_seq_length: int = field(
+        default=384,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    pad_to_max_length: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch (which can "
+            "be faster on GPU but will be slower on TPU)."
+        },
+    )
+    version_2_with_negative: bool = field(
+        default=False, metadata={"help": "If true, some of the examples do not have an answer."}
+    )
+    null_score_diff_threshold: float = field(
+        default=0.0,
+        metadata={
+            "help": "The threshold used to select the null answer: if the best answer has a score that is less than "
+            "the score of the null answer minus this threshold, the null answer is selected for this example. "
+            "Only useful when `version_2_with_negative=True`."
+        },
+    )
+    doc_stride: int = field(
+        default=128,
+        metadata={"help": "When splitting up a long document into chunks, how much stride to take between chunks."},
+    )
+    n_best_size: int = field(
+        default=20,
+        metadata={"help": "The total number of n-best predictions to generate when looking for an answer."},
+    )
+    max_answer_length: int = field(
+        default=30,
+        metadata={
+            "help": "The maximum length of an answer that can be generated. This is needed because the start "
+            "and end predictions are not conditioned on one another."
+        },
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        datasets = load_dataset(extension, data_files=data_files, field="data")
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        use_fast=True,
+    )
+    model = AutoModelForQuestionAnswering.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+    )
+
+    # Tokenizer check: this script requires a fast tokenizer.
+    if not isinstance(tokenizer, PreTrainedTokenizerFast):
+        raise ValueError(
+            "This example script only works for models that have a fast tokenizer. Checkout the big table of models "
+            "at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this "
+            "requirement"
+        )
+
+    # Preprocessing the datasets.
+    # Preprocessing is slighlty different for training and evaluation.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    question_column_name = "question" if "question" in column_names else column_names[0]
+    context_column_name = "context" if "context" in column_names else column_names[1]
+    answer_column_name = "answers" if "answers" in column_names else column_names[2]
+
+    # Padding side determines if we do (question|context) or (context|question).
+    pad_on_right = tokenizer.padding_side == "right"
+
+    # Training preprocessing
+    def prepare_train_features(examples):
+        # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
+        # in one example possible giving several features when a context is long, each of those features having a
+        # context that overlaps a bit the context of the previous feature.
+        tokenized_examples = tokenizer(
+            examples[question_column_name if pad_on_right else context_column_name],
+            examples[context_column_name if pad_on_right else question_column_name],
+            truncation="only_second" if pad_on_right else "only_first",
+            max_length=data_args.max_seq_length,
+            stride=data_args.doc_stride,
+            return_overflowing_tokens=True,
+            return_offsets_mapping=True,
+            padding="max_length" if data_args.pad_to_max_length else False,
+        )
+
+        # Since one example might give us several features if it has a long context, we need a map from a feature to
+        # its corresponding example. This key gives us just that.
+        sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
+        # The offset mappings will give us a map from token to character position in the original context. This will
+        # help us compute the start_positions and end_positions.
+        offset_mapping = tokenized_examples.pop("offset_mapping")
+
+        # Let's label those examples!
+        tokenized_examples["start_positions"] = []
+        tokenized_examples["end_positions"] = []
+
+        for i, offsets in enumerate(offset_mapping):
+            # We will label impossible answers with the index of the CLS token.
+            input_ids = tokenized_examples["input_ids"][i]
+            cls_index = input_ids.index(tokenizer.cls_token_id)
+
+            # Grab the sequence corresponding to that example (to know what is the context and what is the question).
+            sequence_ids = tokenized_examples.sequence_ids(i)
+
+            # One example can give several spans, this is the index of the example containing this span of text.
+            sample_index = sample_mapping[i]
+            answers = examples[answer_column_name][sample_index]
+            # If no answers are given, set the cls_index as answer.
+            if len(answers["answer_start"]) == 0:
+                tokenized_examples["start_positions"].append(cls_index)
+                tokenized_examples["end_positions"].append(cls_index)
+            else:
+                # Start/end character index of the answer in the text.
+                start_char = answers["answer_start"][0]
+                end_char = start_char + len(answers["text"][0])
+
+                # Start token index of the current span in the text.
+                token_start_index = 0
+                while sequence_ids[token_start_index] != (1 if pad_on_right else 0):
+                    token_start_index += 1
+
+                # End token index of the current span in the text.
+                token_end_index = len(input_ids) - 1
+                while sequence_ids[token_end_index] != (1 if pad_on_right else 0):
+                    token_end_index -= 1
+
+                # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
+                if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):
+                    tokenized_examples["start_positions"].append(cls_index)
+                    tokenized_examples["end_positions"].append(cls_index)
+                else:
+                    # Otherwise move the token_start_index and token_end_index to the two ends of the answer.
+                    # Note: we could go after the last offset if the answer is the last word (edge case).
+                    while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:
+                        token_start_index += 1
+                    tokenized_examples["start_positions"].append(token_start_index - 1)
+                    while offsets[token_end_index][1] >= end_char:
+                        token_end_index -= 1
+                    tokenized_examples["end_positions"].append(token_end_index + 1)
+
+        return tokenized_examples
+
+    if training_args.do_train:
+        train_dataset = datasets["train"].map(
+            prepare_train_features,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Validation preprocessing
+    def prepare_validation_features(examples):
+        # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
+        # in one example possible giving several features when a context is long, each of those features having a
+        # context that overlaps a bit the context of the previous feature.
+        tokenized_examples = tokenizer(
+            examples[question_column_name if pad_on_right else context_column_name],
+            examples[context_column_name if pad_on_right else question_column_name],
+            truncation="only_second" if pad_on_right else "only_first",
+            max_length=data_args.max_seq_length,
+            stride=data_args.doc_stride,
+            return_overflowing_tokens=True,
+            return_offsets_mapping=True,
+            padding="max_length" if data_args.pad_to_max_length else False,
+        )
+
+        # Since one example might give us several features if it has a long context, we need a map from a feature to
+        # its corresponding example. This key gives us just that.
+        sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
+
+        # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
+        # corresponding example_id and we will store the offset mappings.
+        tokenized_examples["example_id"] = []
+
+        for i in range(len(tokenized_examples["input_ids"])):
+            # Grab the sequence corresponding to that example (to know what is the context and what is the question).
+            sequence_ids = tokenized_examples.sequence_ids(i)
+            context_index = 1 if pad_on_right else 0
+
+            # One example can give several spans, this is the index of the example containing this span of text.
+            sample_index = sample_mapping[i]
+            tokenized_examples["example_id"].append(examples["id"][sample_index])
+
+            # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
+            # position is part of the context or not.
+            tokenized_examples["offset_mapping"][i] = [
+                (o if sequence_ids[k] == context_index else None)
+                for k, o in enumerate(tokenized_examples["offset_mapping"][i])
+            ]
+
+        return tokenized_examples
+
+    if training_args.do_eval:
+        validation_dataset = datasets["validation"].map(
+            prepare_validation_features,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Data collator
+    # We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
+    # collator.
+    data_collator = default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding(tokenizer)
+
+    # Post-processing:
+    def post_processing_function(examples, features, predictions):
+        # Post-processing: we match the start logits and end logits to answers in the original context.
+        predictions = postprocess_qa_predictions(
+            examples=examples,
+            features=features,
+            predictions=predictions,
+            version_2_with_negative=data_args.version_2_with_negative,
+            n_best_size=data_args.n_best_size,
+            max_answer_length=data_args.max_answer_length,
+            null_score_diff_threshold=data_args.null_score_diff_threshold,
+            output_dir=training_args.output_dir,
+            is_world_process_zero=trainer.is_world_process_zero(),
+        )
+        # Format the result to the format the metric expects.
+        if data_args.version_2_with_negative:
+            formatted_predictions = [
+                {"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in predictions.items()
+            ]
+        else:
+            formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]
+        references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in datasets["validation"]]
+        return EvalPrediction(predictions=formatted_predictions, label_ids=references)
+
+    # TODO: Once the fix lands in a Datasets release, remove the _local here and the squad_v2_local folder.
+    current_dir = os.path.sep.join(os.path.join(__file__).split(os.path.sep)[:-1])
+    metric = load_metric(os.path.join(current_dir, "squad_v2_local") if data_args.version_2_with_negative else "squad")
+
+    def compute_metrics(p: EvalPrediction):
+        return metric.compute(predictions=p.predictions, references=p.label_ids)
+
+    # Initialize our Trainer
+    trainer = QuestionAnsweringTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset if training_args.do_train else None,
+        eval_dataset=validation_dataset if training_args.do_eval else None,
+        eval_examples=datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        post_process_function=post_processing_function,
+        compute_metrics=compute_metrics,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        results = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,512 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning XLNet for question answering with beam search.
+"""
+# You can also adapt this script on your own question answering task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import load_dataset, load_metric
+
+import transformers
+from trainer_qa import QuestionAnsweringTrainer
+from transformers import (
+    DataCollatorWithPadding,
+    EvalPrediction,
+    HfArgumentParser,
+    TrainingArguments,
+    XLNetConfig,
+    XLNetForQuestionAnswering,
+    XLNetTokenizerFast,
+    default_data_collator,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+from utils_qa import postprocess_qa_predictions_with_beam_search
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: Optional[str] = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_seq_length: int = field(
+        default=384,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    pad_to_max_length: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch (which can "
+            "be faster on GPU but will be slower on TPU)."
+        },
+    )
+    version_2_with_negative: bool = field(
+        default=False, metadata={"help": "If true, some of the examples do not have an answer."}
+    )
+    null_score_diff_threshold: float = field(
+        default=0.0,
+        metadata={
+            "help": "The threshold used to select the null answer: if the best answer has a score that is less than "
+            "the score of the null answer minus this threshold, the null answer is selected for this example. "
+            "Only useful when `version_2_with_negative=True`."
+        },
+    )
+    doc_stride: int = field(
+        default=128,
+        metadata={"help": "When splitting up a long document into chunks, how much stride to take between chunks."},
+    )
+    n_best_size: int = field(
+        default=20,
+        metadata={"help": "The total number of n-best predictions to generate when looking for an answer."},
+    )
+    max_answer_length: int = field(
+        default=30,
+        metadata={
+            "help": "The maximum length of an answer that can be generated. This is needed because the start "
+            "and end predictions are not conditioned on one another."
+        },
+    )
+
+    def __post_init__(self):
+        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
+            raise ValueError("Need either a dataset name or a training/validation file.")
+        else:
+            if self.train_file is not None:
+                extension = self.train_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
+            if self.validation_file is not None:
+                extension = self.validation_file.split(".")[-1]
+                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
+    else:
+        data_files = {}
+        if data_args.train_file is not None:
+            data_files["train"] = data_args.train_file
+        if data_args.validation_file is not None:
+            data_files["validation"] = data_args.validation_file
+        extension = data_args.train_file.split(".")[-1]
+        datasets = load_dataset(extension, data_files=data_files, field="data")
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    config = XLNetConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = XLNetTokenizerFast.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    model = XLNetForQuestionAnswering.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+    )
+
+    # Preprocessing the datasets.
+    # Preprocessing is slighlty different for training and evaluation.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    question_column_name = "question" if "question" in column_names else column_names[0]
+    context_column_name = "context" if "context" in column_names else column_names[1]
+    answer_column_name = "answers" if "answers" in column_names else column_names[2]
+
+    # Padding side determines if we do (question|context) or (context|question).
+    pad_on_right = tokenizer.padding_side == "right"
+
+    # Training preprocessing
+    def prepare_train_features(examples):
+        # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
+        # in one example possible giving several features when a context is long, each of those features having a
+        # context that overlaps a bit the context of the previous feature.
+        tokenized_examples = tokenizer(
+            examples[question_column_name if pad_on_right else context_column_name],
+            examples[context_column_name if pad_on_right else question_column_name],
+            truncation="only_second" if pad_on_right else "only_first",
+            max_length=data_args.max_seq_length,
+            stride=data_args.doc_stride,
+            return_overflowing_tokens=True,
+            return_offsets_mapping=True,
+            return_special_tokens_mask=True,
+            return_token_type_ids=True,
+            padding="max_length",
+        )
+
+        # Since one example might give us several features if it has a long context, we need a map from a feature to
+        # its corresponding example. This key gives us just that.
+        sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
+        # The offset mappings will give us a map from token to character position in the original context. This will
+        # help us compute the start_positions and end_positions.
+        offset_mapping = tokenized_examples.pop("offset_mapping")
+        # The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
+        special_tokens = tokenized_examples.pop("special_tokens_mask")
+
+        # Let's label those examples!
+        tokenized_examples["start_positions"] = []
+        tokenized_examples["end_positions"] = []
+        tokenized_examples["is_impossible"] = []
+        tokenized_examples["cls_index"] = []
+        tokenized_examples["p_mask"] = []
+
+        for i, offsets in enumerate(offset_mapping):
+            # We will label impossible answers with the index of the CLS token.
+            input_ids = tokenized_examples["input_ids"][i]
+            cls_index = input_ids.index(tokenizer.cls_token_id)
+            tokenized_examples["cls_index"].append(cls_index)
+
+            # Grab the sequence corresponding to that example (to know what is the context and what is the question).
+            sequence_ids = tokenized_examples["token_type_ids"][i]
+            for k, s in enumerate(special_tokens[i]):
+                if s:
+                    sequence_ids[k] = 3
+            context_idx = 1 if pad_on_right else 0
+
+            # Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
+            # The cls token gets 1.0 too (for predictions of empty answers).
+            tokenized_examples["p_mask"].append(
+                [
+                    0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
+                    for k, s in enumerate(sequence_ids)
+                ]
+            )
+
+            # One example can give several spans, this is the index of the example containing this span of text.
+            sample_index = sample_mapping[i]
+            answers = examples[answer_column_name][sample_index]
+            # If no answers are given, set the cls_index as answer.
+            if len(answers["answer_start"]) == 0:
+                tokenized_examples["start_positions"].append(cls_index)
+                tokenized_examples["end_positions"].append(cls_index)
+                tokenized_examples["is_impossible"].append(1.0)
+            else:
+                # Start/end character index of the answer in the text.
+                start_char = answers["answer_start"][0]
+                end_char = start_char + len(answers["text"][0])
+
+                # Start token index of the current span in the text.
+                token_start_index = 0
+                while sequence_ids[token_start_index] != context_idx:
+                    token_start_index += 1
+
+                # End token index of the current span in the text.
+                token_end_index = len(input_ids) - 1
+                while sequence_ids[token_end_index] != context_idx:
+                    token_end_index -= 1
+                # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
+                if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):
+                    tokenized_examples["start_positions"].append(cls_index)
+                    tokenized_examples["end_positions"].append(cls_index)
+                    tokenized_examples["is_impossible"].append(1.0)
+                else:
+                    # Otherwise move the token_start_index and token_end_index to the two ends of the answer.
+                    # Note: we could go after the last offset if the answer is the last word (edge case).
+                    while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:
+                        token_start_index += 1
+                    tokenized_examples["start_positions"].append(token_start_index - 1)
+                    while offsets[token_end_index][1] >= end_char:
+                        token_end_index -= 1
+                    tokenized_examples["end_positions"].append(token_end_index + 1)
+                    tokenized_examples["is_impossible"].append(0.0)
+
+        return tokenized_examples
+
+    if training_args.do_train:
+        train_dataset = datasets["train"].map(
+            prepare_train_features,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Validation preprocessing
+    def prepare_validation_features(examples):
+        # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
+        # in one example possible giving several features when a context is long, each of those features having a
+        # context that overlaps a bit the context of the previous feature.
+        tokenized_examples = tokenizer(
+            examples[question_column_name if pad_on_right else context_column_name],
+            examples[context_column_name if pad_on_right else question_column_name],
+            truncation="only_second" if pad_on_right else "only_first",
+            max_length=data_args.max_seq_length,
+            stride=data_args.doc_stride,
+            return_overflowing_tokens=True,
+            return_offsets_mapping=True,
+            return_special_tokens_mask=True,
+            return_token_type_ids=True,
+            padding="max_length",
+        )
+
+        # Since one example might give us several features if it has a long context, we need a map from a feature to
+        # its corresponding example. This key gives us just that.
+        sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
+
+        # The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
+        special_tokens = tokenized_examples.pop("special_tokens_mask")
+
+        # For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
+        # corresponding example_id and we will store the offset mappings.
+        tokenized_examples["example_id"] = []
+
+        # We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
+        tokenized_examples["cls_index"] = []
+        tokenized_examples["p_mask"] = []
+
+        for i, input_ids in enumerate(tokenized_examples["input_ids"]):
+            # Find the CLS token in the input ids.
+            cls_index = input_ids.index(tokenizer.cls_token_id)
+            tokenized_examples["cls_index"].append(cls_index)
+
+            # Grab the sequence corresponding to that example (to know what is the context and what is the question).
+            sequence_ids = tokenized_examples["token_type_ids"][i]
+            for k, s in enumerate(special_tokens[i]):
+                if s:
+                    sequence_ids[k] = 3
+            context_idx = 1 if pad_on_right else 0
+
+            # Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
+            tokenized_examples["p_mask"].append(
+                [
+                    0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
+                    for k, s in enumerate(sequence_ids)
+                ]
+            )
+
+            # One example can give several spans, this is the index of the example containing this span of text.
+            sample_index = sample_mapping[i]
+            tokenized_examples["example_id"].append(examples["id"][sample_index])
+
+            # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
+            # position is part of the context or not.
+            tokenized_examples["offset_mapping"][i] = [
+                (o if sequence_ids[k] == context_idx else None)
+                for k, o in enumerate(tokenized_examples["offset_mapping"][i])
+            ]
+
+        return tokenized_examples
+
+    if training_args.do_eval:
+        validation_dataset = datasets["validation"].map(
+            prepare_validation_features,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=column_names,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+    # Data collator
+    # We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
+    # collator.
+    data_collator = default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding(tokenizer)
+
+    # Post-processing:
+    def post_processing_function(examples, features, predictions):
+        # Post-processing: we match the start logits and end logits to answers in the original context.
+        predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
+            examples=examples,
+            features=features,
+            predictions=predictions,
+            version_2_with_negative=data_args.version_2_with_negative,
+            n_best_size=data_args.n_best_size,
+            max_answer_length=data_args.max_answer_length,
+            start_n_top=model.config.start_n_top,
+            end_n_top=model.config.end_n_top,
+            output_dir=training_args.output_dir,
+            is_world_process_zero=trainer.is_world_process_zero(),
+        )
+        # Format the result to the format the metric expects.
+        if data_args.version_2_with_negative:
+            formatted_predictions = [
+                {"id": k, "prediction_text": v, "no_answer_probability": scores_diff_json[k]}
+                for k, v in predictions.items()
+            ]
+        else:
+            formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]
+        references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in datasets["validation"]]
+        return EvalPrediction(predictions=formatted_predictions, label_ids=references)
+
+    # TODO: Once the fix lands in a Datasets release, remove the _local here and the squad_v2_local folder.
+    current_dir = os.path.sep.join(os.path.join(__file__).split(os.path.sep)[:-1])
+    metric = load_metric(os.path.join(current_dir, "squad_v2_local") if data_args.version_2_with_negative else "squad")
+
+    def compute_metrics(p: EvalPrediction):
+        return metric.compute(predictions=p.predictions, references=p.label_ids)
+
+    # Initialize our Trainer
+    trainer = QuestionAnsweringTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset if training_args.do_train else None,
+        eval_dataset=validation_dataset if training_args.do_eval else None,
+        eval_examples=datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        post_process_function=post_processing_function,
+        compute_metrics=compute_metrics,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        results = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,322 @@
+"""Official evaluation script for SQuAD version 2.0.
+
+In addition to basic functionality, we also compute additional statistics and
+plot precision-recall curves if an additional na_prob.json file is provided.
+This file is expected to map question ID's to the model's predicted probability
+that a question is unanswerable.
+"""
+import argparse
+import collections
+import json
+import os
+import re
+import string
+import sys
+
+import numpy as np
+
+
+OPTS = None
+
+
+def parse_args():
+    parser = argparse.ArgumentParser("Official evaluation script for SQuAD version 2.0.")
+    parser.add_argument("data_file", metavar="data.json", help="Input data JSON file.")
+    parser.add_argument("pred_file", metavar="pred.json", help="Model predictions.")
+    parser.add_argument(
+        "--out-file", "-o", metavar="eval.json", help="Write accuracy metrics to file (default is stdout)."
+    )
+    parser.add_argument(
+        "--na-prob-file", "-n", metavar="na_prob.json", help="Model estimates of probability of no answer."
+    )
+    parser.add_argument(
+        "--na-prob-thresh",
+        "-t",
+        type=float,
+        default=1.0,
+        help='Predict "" if no-answer probability exceeds this (default = 1.0).',
+    )
+    parser.add_argument(
+        "--out-image-dir", "-p", metavar="out_images", default=None, help="Save precision-recall curves to directory."
+    )
+    parser.add_argument("--verbose", "-v", action="store_true")
+    if len(sys.argv) == 1:
+        parser.print_help()
+        sys.exit(1)
+    return parser.parse_args()
+
+
+def make_qid_to_has_ans(dataset):
+    qid_to_has_ans = {}
+    for article in dataset:
+        for p in article["paragraphs"]:
+            for qa in p["qas"]:
+                qid_to_has_ans[qa["id"]] = bool(qa["answers"]["text"])
+    return qid_to_has_ans
+
+
+def normalize_answer(s):
+    """Lower text and remove punctuation, articles and extra whitespace."""
+
+    def remove_articles(text):
+        regex = re.compile(r"\b(a|an|the)\b", re.UNICODE)
+        return re.sub(regex, " ", text)
+
+    def white_space_fix(text):
+        return " ".join(text.split())
+
+    def remove_punc(text):
+        exclude = set(string.punctuation)
+        return "".join(ch for ch in text if ch not in exclude)
+
+    def lower(text):
+        return text.lower()
+
+    return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+
+def get_tokens(s):
+    if not s:
+        return []
+    return normalize_answer(s).split()
+
+
+def compute_exact(a_gold, a_pred):
+    return int(normalize_answer(a_gold) == normalize_answer(a_pred))
+
+
+def compute_f1(a_gold, a_pred):
+    gold_toks = get_tokens(a_gold)
+    pred_toks = get_tokens(a_pred)
+    common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
+    num_same = sum(common.values())
+    if len(gold_toks) == 0 or len(pred_toks) == 0:
+        # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
+        return int(gold_toks == pred_toks)
+    if num_same == 0:
+        return 0
+    precision = 1.0 * num_same / len(pred_toks)
+    recall = 1.0 * num_same / len(gold_toks)
+    f1 = (2 * precision * recall) / (precision + recall)
+    return f1
+
+
+def get_raw_scores(dataset, preds):
+    exact_scores = {}
+    f1_scores = {}
+    for article in dataset:
+        for p in article["paragraphs"]:
+            for qa in p["qas"]:
+                qid = qa["id"]
+                gold_answers = [t for t in qa["answers"]["text"] if normalize_answer(t)]
+                if not gold_answers:
+                    # For unanswerable questions, only correct answer is empty string
+                    gold_answers = [""]
+                if qid not in preds:
+                    print("Missing prediction for %s" % qid)
+                    continue
+                a_pred = preds[qid]
+                # Take max over all gold answers
+                exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
+                f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
+    return exact_scores, f1_scores
+
+
+def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
+    new_scores = {}
+    for qid, s in scores.items():
+        pred_na = na_probs[qid] > na_prob_thresh
+        if pred_na:
+            new_scores[qid] = float(not qid_to_has_ans[qid])
+        else:
+            new_scores[qid] = s
+    return new_scores
+
+
+def make_eval_dict(exact_scores, f1_scores, qid_list=None):
+    if not qid_list:
+        total = len(exact_scores)
+        return collections.OrderedDict(
+            [
+                ("exact", 100.0 * sum(exact_scores.values()) / total),
+                ("f1", 100.0 * sum(f1_scores.values()) / total),
+                ("total", total),
+            ]
+        )
+    else:
+        total = len(qid_list)
+        return collections.OrderedDict(
+            [
+                ("exact", 100.0 * sum(exact_scores[k] for k in qid_list) / total),
+                ("f1", 100.0 * sum(f1_scores[k] for k in qid_list) / total),
+                ("total", total),
+            ]
+        )
+
+
+def merge_eval(main_eval, new_eval, prefix):
+    for k in new_eval:
+        main_eval["%s_%s" % (prefix, k)] = new_eval[k]
+
+
+def plot_pr_curve(precisions, recalls, out_image, title):
+    plt.step(recalls, precisions, color="b", alpha=0.2, where="post")
+    plt.fill_between(recalls, precisions, step="post", alpha=0.2, color="b")
+    plt.xlabel("Recall")
+    plt.ylabel("Precision")
+    plt.xlim([0.0, 1.05])
+    plt.ylim([0.0, 1.05])
+    plt.title(title)
+    plt.savefig(out_image)
+    plt.clf()
+
+
+def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans, out_image=None, title=None):
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    true_pos = 0.0
+    cur_p = 1.0
+    cur_r = 0.0
+    precisions = [1.0]
+    recalls = [0.0]
+    avg_prec = 0.0
+    for i, qid in enumerate(qid_list):
+        if qid_to_has_ans[qid]:
+            true_pos += scores[qid]
+        cur_p = true_pos / float(i + 1)
+        cur_r = true_pos / float(num_true_pos)
+        if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i + 1]]:
+            # i.e., if we can put a threshold after this point
+            avg_prec += cur_p * (cur_r - recalls[-1])
+            precisions.append(cur_p)
+            recalls.append(cur_r)
+    if out_image:
+        plot_pr_curve(precisions, recalls, out_image, title)
+    return {"ap": 100.0 * avg_prec}
+
+
+def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs, qid_to_has_ans, out_image_dir):
+    if out_image_dir and not os.path.exists(out_image_dir):
+        os.makedirs(out_image_dir)
+    num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
+    if num_true_pos == 0:
+        return
+    pr_exact = make_precision_recall_eval(
+        exact_raw,
+        na_probs,
+        num_true_pos,
+        qid_to_has_ans,
+        out_image=os.path.join(out_image_dir, "pr_exact.png"),
+        title="Precision-Recall curve for Exact Match score",
+    )
+    pr_f1 = make_precision_recall_eval(
+        f1_raw,
+        na_probs,
+        num_true_pos,
+        qid_to_has_ans,
+        out_image=os.path.join(out_image_dir, "pr_f1.png"),
+        title="Precision-Recall curve for F1 score",
+    )
+    oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
+    pr_oracle = make_precision_recall_eval(
+        oracle_scores,
+        na_probs,
+        num_true_pos,
+        qid_to_has_ans,
+        out_image=os.path.join(out_image_dir, "pr_oracle.png"),
+        title="Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)",
+    )
+    merge_eval(main_eval, pr_exact, "pr_exact")
+    merge_eval(main_eval, pr_f1, "pr_f1")
+    merge_eval(main_eval, pr_oracle, "pr_oracle")
+
+
+def histogram_na_prob(na_probs, qid_list, image_dir, name):
+    if not qid_list:
+        return
+    x = [na_probs[k] for k in qid_list]
+    weights = np.ones_like(x) / float(len(x))
+    plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
+    plt.xlabel("Model probability of no-answer")
+    plt.ylabel("Proportion of dataset")
+    plt.title("Histogram of no-answer probability: %s" % name)
+    plt.savefig(os.path.join(image_dir, "na_prob_hist_%s.png" % name))
+    plt.clf()
+
+
+def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
+    num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+    cur_score = num_no_ans
+    best_score = cur_score
+    best_thresh = 0.0
+    qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+    for i, qid in enumerate(qid_list):
+        if qid not in scores:
+            continue
+        if qid_to_has_ans[qid]:
+            diff = scores[qid]
+        else:
+            if preds[qid]:
+                diff = -1
+            else:
+                diff = 0
+        cur_score += diff
+        if cur_score > best_score:
+            best_score = cur_score
+            best_thresh = na_probs[qid]
+    return 100.0 * best_score / len(scores), best_thresh
+
+
+def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+    best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
+    best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
+    main_eval["best_exact"] = best_exact
+    main_eval["best_exact_thresh"] = exact_thresh
+    main_eval["best_f1"] = best_f1
+    main_eval["best_f1_thresh"] = f1_thresh
+
+
+def main():
+    with open(OPTS.data_file) as f:
+        dataset_json = json.load(f)
+        dataset = dataset_json["data"]
+    with open(OPTS.pred_file) as f:
+        preds = json.load(f)
+    if OPTS.na_prob_file:
+        with open(OPTS.na_prob_file) as f:
+            na_probs = json.load(f)
+    else:
+        na_probs = {k: 0.0 for k in preds}
+    qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
+    has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+    no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+    exact_raw, f1_raw = get_raw_scores(dataset, preds)
+    exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans, OPTS.na_prob_thresh)
+    f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans, OPTS.na_prob_thresh)
+    out_eval = make_eval_dict(exact_thresh, f1_thresh)
+    if has_ans_qids:
+        has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
+        merge_eval(out_eval, has_ans_eval, "HasAns")
+    if no_ans_qids:
+        no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
+        merge_eval(out_eval, no_ans_eval, "NoAns")
+    if OPTS.na_prob_file:
+        find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
+    if OPTS.na_prob_file and OPTS.out_image_dir:
+        run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs, qid_to_has_ans, OPTS.out_image_dir)
+        histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, "hasAns")
+        histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, "noAns")
+    if OPTS.out_file:
+        with open(OPTS.out_file, "w") as f:
+            json.dump(out_eval, f)
+    else:
+        print(json.dumps(out_eval, indent=2))
+
+
+if __name__ == "__main__":
+    OPTS = parse_args()
+    if OPTS.out_image_dir:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        import matplotlib.pyplot as plt
+    main()
@@ -0,0 +1,128 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Datasets Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" SQuAD v2 metric. """
+
+import datasets
+
+from .evaluate import (
+    apply_no_ans_threshold,
+    find_all_best_thresh,
+    get_raw_scores,
+    make_eval_dict,
+    make_qid_to_has_ans,
+    merge_eval,
+)
+
+
+_CITATION = """\
+@inproceedings{Rajpurkar2016SQuAD10,
+  title={SQuAD: 100, 000+ Questions for Machine Comprehension of Text},
+  author={Pranav Rajpurkar and Jian Zhang and Konstantin Lopyrev and Percy Liang},
+  booktitle={EMNLP},
+  year={2016}
+}
+"""
+
+_DESCRIPTION = """
+This metric wrap the official scoring script for version 2 of the Stanford Question
+Answering Dataset (SQuAD).
+
+Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by
+crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span,
+from the corresponding reading passage, or the question might be unanswerable.
+
+SQuAD2.0 combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions
+written adversarially by crowdworkers to look similar to answerable ones.
+To do well on SQuAD2.0, systems must not only answer questions when possible, but also
+determine when no answer is supported by the paragraph and abstain from answering.
+"""
+
+_KWARGS_DESCRIPTION = """
+Computes SQuAD v2 scores (F1 and EM).
+Args:
+    predictions: List of triple for question-answers to score with the following elements:
+        - the question-answer 'id' field as given in the references (see below)
+        - the text of the answer
+        - the probability that the question has no answer
+    references: List of question-answers dictionaries with the following key-values:
+            - 'id': id of the question-answer pair (see above),
+            - 'answers': a list of Dict {'text': text of the answer as a string}
+    no_answer_threshold: float
+        Probability threshold to decide that a question has no answer.
+Returns:
+    'exact': Exact match (the normalized answer exactly match the gold answer)
+    'f1': The F-score of predicted tokens versus the gold answer
+    'total': Number of score considered
+    'HasAns_exact': Exact match (the normalized answer exactly match the gold answer)
+    'HasAns_f1': The F-score of predicted tokens versus the gold answer
+    'HasAns_total': Number of score considered
+    'NoAns_exact': Exact match (the normalized answer exactly match the gold answer)
+    'NoAns_f1': The F-score of predicted tokens versus the gold answer
+    'NoAns_total': Number of score considered
+    'best_exact': Best exact match (with varying threshold)
+    'best_exact_thresh': No-answer probability threshold associated to the best exact match
+    'best_f1': Best F1 (with varying threshold)
+    'best_f1_thresh': No-answer probability threshold associated to the best F1
+"""
+
+
+class SquadV2(datasets.Metric):
+    def _info(self):
+        return datasets.MetricInfo(
+            description=_DESCRIPTION,
+            citation=_CITATION,
+            inputs_description=_KWARGS_DESCRIPTION,
+            features=datasets.Features(
+                {
+                    "predictions": {
+                        "id": datasets.Value("string"),
+                        "prediction_text": datasets.Value("string"),
+                        "no_answer_probability": datasets.Value("float32"),
+                    },
+                    "references": {
+                        "id": datasets.Value("string"),
+                        "answers": datasets.features.Sequence(
+                            {"text": datasets.Value("string"), "answer_start": datasets.Value("int32")}
+                        ),
+                    },
+                }
+            ),
+            codebase_urls=["https://rajpurkar.github.io/SQuAD-explorer/"],
+            reference_urls=["https://rajpurkar.github.io/SQuAD-explorer/"],
+        )
+
+    def _compute(self, predictions, references, no_answer_threshold=1.0):
+        no_answer_probabilities = dict((p["id"], p["no_answer_probability"]) for p in predictions)
+        dataset = [{"paragraphs": [{"qas": references}]}]
+        predictions = dict((p["id"], p["prediction_text"]) for p in predictions)
+
+        qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
+        has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+        no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+
+        exact_raw, f1_raw = get_raw_scores(dataset, predictions)
+        exact_thresh = apply_no_ans_threshold(exact_raw, no_answer_probabilities, qid_to_has_ans, no_answer_threshold)
+        f1_thresh = apply_no_ans_threshold(f1_raw, no_answer_probabilities, qid_to_has_ans, no_answer_threshold)
+        out_eval = make_eval_dict(exact_thresh, f1_thresh)
+
+        if has_ans_qids:
+            has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
+            merge_eval(out_eval, has_ans_eval, "HasAns")
+        if no_ans_qids:
+            no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
+            merge_eval(out_eval, no_ans_eval, "NoAns")
+        find_all_best_thresh(out_eval, predictions, exact_raw, f1_raw, no_answer_probabilities, qid_to_has_ans)
+
+        return out_eval
@@ -0,0 +1,104 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+A subclass of `Trainer` specific to Question-Answering tasks
+"""
+
+from transformers import Trainer, is_datasets_available, is_torch_tpu_available
+from transformers.trainer_utils import PredictionOutput
+
+
+if is_datasets_available():
+    import datasets
+
+if is_torch_tpu_available():
+    import torch_xla.core.xla_model as xm
+    import torch_xla.debug.metrics as met
+
+
+class QuestionAnsweringTrainer(Trainer):
+    def __init__(self, *args, eval_examples=None, post_process_function=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.eval_examples = eval_examples
+        self.post_process_function = post_process_function
+
+    def evaluate(self, eval_dataset=None, eval_examples=None, ignore_keys=None):
+        eval_dataset = self.eval_dataset if eval_dataset is None else eval_dataset
+        eval_dataloader = self.get_eval_dataloader(eval_dataset)
+        eval_examples = self.eval_examples if eval_examples is None else eval_examples
+
+        # Temporarily disable metric computation, we will do it in the loop here.
+        compute_metrics = self.compute_metrics
+        self.compute_metrics = None
+        try:
+            output = self.prediction_loop(
+                eval_dataloader,
+                description="Evaluation",
+                # No point gathering the predictions if there are no metrics, otherwise we defer to
+                # self.args.prediction_loss_only
+                prediction_loss_only=True if compute_metrics is None else None,
+                ignore_keys=ignore_keys,
+            )
+        finally:
+            self.compute_metrics = compute_metrics
+
+        # We might have removed columns from the dataset so we put them back.
+        if isinstance(eval_dataset, datasets.Dataset):
+            eval_dataset.set_format(type=eval_dataset.format["type"], columns=list(eval_dataset.features.keys()))
+
+        if self.post_process_function is not None and self.compute_metrics is not None:
+            eval_preds = self.post_process_function(eval_examples, eval_dataset, output.predictions)
+            metrics = self.compute_metrics(eval_preds)
+
+            self.log(metrics)
+        else:
+            metrics = {}
+
+        if self.args.tpu_metrics_debug or self.args.debug:
+            # tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)
+            xm.master_print(met.metrics_report())
+
+        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, metrics)
+        return metrics
+
+    def predict(self, test_dataset, test_examples, ignore_keys=None):
+        test_dataloader = self.get_test_dataloader(test_dataset)
+
+        # Temporarily disable metric computation, we will do it in the loop here.
+        compute_metrics = self.compute_metrics
+        self.compute_metrics = None
+        try:
+            output = self.prediction_loop(
+                test_dataloader,
+                description="Evaluation",
+                # No point gathering the predictions if there are no metrics, otherwise we defer to
+                # self.args.prediction_loss_only
+                prediction_loss_only=True if compute_metrics is None else None,
+                ignore_keys=ignore_keys,
+            )
+        finally:
+            self.compute_metrics = compute_metrics
+
+        if self.post_process_function is None or self.compute_metrics is None:
+            return output
+
+        # We might have removed columns from the dataset so we put them back.
+        if isinstance(test_dataset, datasets.Dataset):
+            test_dataset.set_format(type=test_dataset.format["type"], columns=list(test_dataset.features.keys()))
+
+        eval_preds = self.post_process_function(test_examples, test_dataset, output.predictions)
+        metrics = self.compute_metrics(eval_preds)
+
+        return PredictionOutput(predictions=eval_preds.predictions, label_ids=eval_preds.label_ids, metrics=metrics)
@@ -0,0 +1,429 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Post-processing utilities for question answering.
+"""
+import collections
+import json
+import logging
+import os
+from typing import Optional, Tuple
+
+import numpy as np
+from tqdm.auto import tqdm
+
+
+logger = logging.getLogger(__name__)
+
+
+def postprocess_qa_predictions(
+    examples,
+    features,
+    predictions: Tuple[np.ndarray, np.ndarray],
+    version_2_with_negative: bool = False,
+    n_best_size: int = 20,
+    max_answer_length: int = 30,
+    null_score_diff_threshold: float = 0.0,
+    output_dir: Optional[str] = None,
+    prefix: Optional[str] = None,
+    is_world_process_zero: bool = True,
+):
+    """
+    Post-processes the predictions of a question-answering model to convert them to answers that are substrings of the
+    original contexts. This is the base postprocessing functions for models that only return start and end logits.
+
+    Args:
+        examples: The non-preprocessed dataset (see the main script for more information).
+        features: The processed dataset (see the main script for more information).
+        predictions (:obj:`Tuple[np.ndarray, np.ndarray]`):
+            The predictions of the model: two arrays containing the start logits and the end logits respectively. Its
+            first dimension must match the number of elements of :obj:`features`.
+        version_2_with_negative (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the underlying dataset contains examples with no answers.
+        n_best_size (:obj:`int`, `optional`, defaults to 20):
+            The total number of n-best predictions to generate when looking for an answer.
+        max_answer_length (:obj:`int`, `optional`, defaults to 30):
+            The maximum length of an answer that can be generated. This is needed because the start and end predictions
+            are not conditioned on one another.
+        null_score_diff_threshold (:obj:`float`, `optional`, defaults to 0):
+            The threshold used to select the null answer: if the best answer has a score that is less than the score of
+            the null answer minus this threshold, the null answer is selected for this example (note that the score of
+            the null answer for an example giving several features is the minimum of the scores for the null answer on
+            each feature: all features must be aligned on the fact they `want` to predict a null answer).
+
+            Only useful when :obj:`version_2_with_negative` is :obj:`True`.
+        output_dir (:obj:`str`, `optional`):
+            If provided, the dictionaries of predictions, n_best predictions (with their scores and logits) and, if
+            :obj:`version_2_with_negative=True`, the dictionary of the scores differences between best and null
+            answers, are saved in `output_dir`.
+        prefix (:obj:`str`, `optional`):
+            If provided, the dictionaries mentioned above are saved with `prefix` added to their names.
+        is_world_process_zero (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this process is the main process or not (used to determine if logging/saves should be done).
+    """
+    assert len(predictions) == 2, "`predictions` should be a tuple with two elements (start_logits, end_logits)."
+    all_start_logits, all_end_logits = predictions
+
+    assert len(predictions[0]) == len(
+        features
+    ), f"Got {len(predictions[0])} predicitions and {len(features)} features."
+
+    # Build a map example to its corresponding features.
+    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
+    features_per_example = collections.defaultdict(list)
+    for i, feature in enumerate(features):
+        features_per_example[example_id_to_index[feature["example_id"]]].append(i)
+
+    # The dictionaries we have to fill.
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    if version_2_with_negative:
+        scores_diff_json = collections.OrderedDict()
+
+    # Logging.
+    logger.setLevel(logging.INFO if is_world_process_zero else logging.WARN)
+    logger.info(f"Post-processing {len(examples)} example predictions split into {len(features)} features.")
+
+    # Let's loop over all the examples!
+    for example_index, example in enumerate(tqdm(examples)):
+        # Those are the indices of the features associated to the current example.
+        feature_indices = features_per_example[example_index]
+
+        min_null_prediction = None
+        prelim_predictions = []
+
+        # Looping through all the features associated to the current example.
+        for feature_index in feature_indices:
+            # We grab the predictions of the model for this feature.
+            start_logits = all_start_logits[feature_index]
+            end_logits = all_end_logits[feature_index]
+            # This is what will allow us to map some the positions in our logits to span of texts in the original
+            # context.
+            offset_mapping = features[feature_index]["offset_mapping"]
+            # Optional `token_is_max_context`, if provided we will remove answers that do not have the maximum context
+            # available in the current feature.
+            token_is_max_context = features[feature_index].get("token_is_max_context", None)
+
+            # Update minimum null prediction.
+            feature_null_score = start_logits[0] + end_logits[0]
+            if min_null_prediction is None or min_null_prediction["score"] < feature_null_score:
+                min_null_prediction = {
+                    "offsets": (0, 0),
+                    "score": feature_null_score,
+                    "start_logit": start_logits[0],
+                    "end_logit": end_logits[0],
+                }
+
+            # Go through all possibilities for the `n_best_size` greater start and end logits.
+            start_indexes = np.argsort(start_logits)[-1 : -n_best_size - 1 : -1].tolist()
+            end_indexes = np.argsort(end_logits)[-1 : -n_best_size - 1 : -1].tolist()
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # Don't consider out-of-scope answers, either because the indices are out of bounds or correspond
+                    # to part of the input_ids that are not in the context.
+                    if (
+                        start_index >= len(offset_mapping)
+                        or end_index >= len(offset_mapping)
+                        or offset_mapping[start_index] is None
+                        or offset_mapping[end_index] is None
+                    ):
+                        continue
+                    # Don't consider answers with a length that is either < 0 or > max_answer_length.
+                    if end_index < start_index or end_index - start_index + 1 > max_answer_length:
+                        continue
+                    # Don't consider answer that don't have the maximum context available (if such information is
+                    # provided).
+                    if token_is_max_context is not None and not token_is_max_context.get(str(start_index), False):
+                        continue
+                    prelim_predictions.append(
+                        {
+                            "offsets": (offset_mapping[start_index][0], offset_mapping[end_index][1]),
+                            "score": start_logits[start_index] + end_logits[end_index],
+                            "start_logit": start_logits[start_index],
+                            "end_logit": end_logits[end_index],
+                        }
+                    )
+        if version_2_with_negative:
+            # Add the minimum null prediction
+            prelim_predictions.append(min_null_prediction)
+            null_score = min_null_prediction["score"]
+
+        # Only keep the best `n_best_size` predictions.
+        predictions = sorted(prelim_predictions, key=lambda x: x["score"], reverse=True)[:n_best_size]
+
+        # Add back the minimum null prediction if it was removed because of its low score.
+        if version_2_with_negative and not any(p["offsets"] == (0, 0) for p in predictions):
+            predictions.append(min_null_prediction)
+
+        # Use the offsets to gather the answer text in the original context.
+        context = example["context"]
+        for pred in predictions:
+            offsets = pred.pop("offsets")
+            pred["text"] = context[offsets[0] : offsets[1]]
+
+        # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid
+        # failure.
+        if len(predictions) == 0 or (len(predictions) == 1 and predictions[0]["text"] == ""):
+            predictions.insert(0, {"text": "empty", "start_logit": 0.0, "end_logit": 0.0, "score": 0.0})
+
+        # Compute the softmax of all scores (we do it with numpy to stay independent from torch/tf in this file, using
+        # the LogSumExp trick).
+        scores = np.array([pred.pop("score") for pred in predictions])
+        exp_scores = np.exp(scores - np.max(scores))
+        probs = exp_scores / exp_scores.sum()
+
+        # Include the probabilities in our predictions.
+        for prob, pred in zip(probs, predictions):
+            pred["probability"] = prob
+
+        # Pick the best prediction. If the null answer is not possible, this is easy.
+        if not version_2_with_negative:
+            all_predictions[example["id"]] = predictions[0]["text"]
+        else:
+            # Otherwise we first need to find the best non-empty prediction.
+            i = 0
+            while predictions[i]["text"] == "":
+                i += 1
+            best_non_null_pred = predictions[i]
+
+            # Then we compare to the null prediction using the threshold.
+            score_diff = null_score - best_non_null_pred["start_logit"] - best_non_null_pred["end_logit"]
+            scores_diff_json[example["id"]] = float(score_diff)  # To be JSON-serializable.
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example["id"]] = ""
+            else:
+                all_predictions[example["id"]] = best_non_null_pred["text"]
+
+        # Make `predictions` JSON-serializable by casting np.float back to float.
+        all_nbest_json[example["id"]] = [
+            {k: (float(v) if isinstance(v, (np.float32, np.float64)) else v) for k, v in pred.items()}
+            for pred in predictions
+        ]
+
+    # If we have an output_dir, let's save all those dicts.
+    if output_dir is not None:
+        assert os.path.isdir(output_dir), f"{output_dir} is not a directory."
+
+        prediction_file = os.path.join(
+            output_dir, "predictions.json" if prefix is None else f"predictions_{prefix}".json
+        )
+        nbest_file = os.path.join(
+            output_dir, "nbest_predictions.json" if prefix is None else f"nbest_predictions_{prefix}".json
+        )
+        if version_2_with_negative:
+            null_odds_file = os.path.join(
+                output_dir, "null_odds.json" if prefix is None else f"null_odds_{prefix}".json
+            )
+
+        logger.info(f"Saving predictions to {prediction_file}.")
+        with open(prediction_file, "w") as writer:
+            writer.write(json.dumps(all_predictions, indent=4) + "\n")
+        logger.info(f"Saving nbest_preds to {nbest_file}.")
+        with open(nbest_file, "w") as writer:
+            writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+        if version_2_with_negative:
+            logger.info(f"Saving null_odds to {null_odds_file}.")
+            with open(null_odds_file, "w") as writer:
+                writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+def postprocess_qa_predictions_with_beam_search(
+    examples,
+    features,
+    predictions: Tuple[np.ndarray, np.ndarray],
+    version_2_with_negative: bool = False,
+    n_best_size: int = 20,
+    max_answer_length: int = 30,
+    start_n_top: int = 5,
+    end_n_top: int = 5,
+    output_dir: Optional[str] = None,
+    prefix: Optional[str] = None,
+    is_world_process_zero: bool = True,
+):
+    """
+    Post-processes the predictions of a question-answering model with beam search to convert them to answers that are substrings of the
+    original contexts. This is the postprocessing functions for models that return start and end logits, indices, as well as
+    cls token predictions.
+
+    Args:
+        examples: The non-preprocessed dataset (see the main script for more information).
+        features: The processed dataset (see the main script for more information).
+        predictions (:obj:`Tuple[np.ndarray, np.ndarray]`):
+            The predictions of the model: two arrays containing the start logits and the end logits respectively. Its
+            first dimension must match the number of elements of :obj:`features`.
+        version_2_with_negative (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the underlying dataset contains examples with no answers.
+        n_best_size (:obj:`int`, `optional`, defaults to 20):
+            The total number of n-best predictions to generate when looking for an answer.
+        max_answer_length (:obj:`int`, `optional`, defaults to 30):
+            The maximum length of an answer that can be generated. This is needed because the start and end predictions
+            are not conditioned on one another.
+        start_n_top (:obj:`int`, `optional`, defaults to 5):
+            The number of top start logits too keep when searching for the :obj:`n_best_size` predictions.
+        end_n_top (:obj:`int`, `optional`, defaults to 5):
+            The number of top end logits too keep when searching for the :obj:`n_best_size` predictions.
+        output_dir (:obj:`str`, `optional`):
+            If provided, the dictionaries of predictions, n_best predictions (with their scores and logits) and, if
+            :obj:`version_2_with_negative=True`, the dictionary of the scores differences between best and null
+            answers, are saved in `output_dir`.
+        prefix (:obj:`str`, `optional`):
+            If provided, the dictionaries mentioned above are saved with `prefix` added to their names.
+        is_world_process_zero (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether this process is the main process or not (used to determine if logging/saves should be done).
+    """
+    assert len(predictions) == 5, "`predictions` should be a tuple with five elements."
+    start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits = predictions
+
+    assert len(predictions[0]) == len(
+        features
+    ), f"Got {len(predictions[0])} predicitions and {len(features)} features."
+
+    # Build a map example to its corresponding features.
+    example_id_to_index = {k: i for i, k in enumerate(examples["id"])}
+    features_per_example = collections.defaultdict(list)
+    for i, feature in enumerate(features):
+        features_per_example[example_id_to_index[feature["example_id"]]].append(i)
+
+    # The dictionaries we have to fill.
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict() if version_2_with_negative else None
+
+    # Logging.
+    logger.setLevel(logging.INFO if is_world_process_zero else logging.WARN)
+    logger.info(f"Post-processing {len(examples)} example predictions split into {len(features)} features.")
+
+    # Let's loop over all the examples!
+    for example_index, example in enumerate(tqdm(examples)):
+        # Those are the indices of the features associated to the current example.
+        feature_indices = features_per_example[example_index]
+
+        min_null_score = None
+        prelim_predictions = []
+
+        # Looping through all the features associated to the current example.
+        for feature_index in feature_indices:
+            # We grab the predictions of the model for this feature.
+            start_log_prob = start_top_log_probs[feature_index]
+            start_indexes = start_top_index[feature_index]
+            end_log_prob = end_top_log_probs[feature_index]
+            end_indexes = end_top_index[feature_index]
+            feature_null_score = cls_logits[feature_index]
+            # This is what will allow us to map some the positions in our logits to span of texts in the original
+            # context.
+            offset_mapping = features[feature_index]["offset_mapping"]
+            # Optional `token_is_max_context`, if provided we will remove answers that do not have the maximum context
+            # available in the current feature.
+            token_is_max_context = features[feature_index].get("token_is_max_context", None)
+
+            # Update minimum null prediction
+            if min_null_score is None or feature_null_score < min_null_score:
+                min_null_score = feature_null_score
+
+            # Go through all possibilities for the `n_start_top`/`n_end_top` greater start and end logits.
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_index = start_indexes[i]
+                    j_index = i * end_n_top + j
+                    end_index = end_indexes[j_index]
+                    # Don't consider out-of-scope answers (last part of the test should be unnecessary because of the
+                    # p_mask but let's not take any risk)
+                    if (
+                        start_index >= len(offset_mapping)
+                        or end_index >= len(offset_mapping)
+                        or offset_mapping[start_index] is None
+                        or offset_mapping[end_index] is None
+                    ):
+                        continue
+                    # Don't consider answers with a length negative or > max_answer_length.
+                    if end_index < start_index or end_index - start_index + 1 > max_answer_length:
+                        continue
+                    # Don't consider answer that don't have the maximum context available (if such information is
+                    # provided).
+                    if token_is_max_context is not None and not token_is_max_context.get(str(start_index), False):
+                        continue
+                    prelim_predictions.append(
+                        {
+                            "offsets": (offset_mapping[start_index][0], offset_mapping[end_index][1]),
+                            "score": start_log_prob[i] + end_log_prob[j_index],
+                            "start_log_prob": start_log_prob[i],
+                            "end_log_prob": end_log_prob[j_index],
+                        }
+                    )
+
+        # Only keep the best `n_best_size` predictions.
+        predictions = sorted(prelim_predictions, key=lambda x: x["score"], reverse=True)[:n_best_size]
+
+        # Use the offsets to gather the answer text in the original context.
+        context = example["context"]
+        for pred in predictions:
+            offsets = pred.pop("offsets")
+            pred["text"] = context[offsets[0] : offsets[1]]
+
+        # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid
+        # failure.
+        if len(predictions) == 0:
+            predictions.insert(0, {"text": "", "start_logit": -1e-6, "end_logit": -1e-6, "score": -2e-6})
+
+        # Compute the softmax of all scores (we do it with numpy to stay independent from torch/tf in this file, using
+        # the LogSumExp trick).
+        scores = np.array([pred.pop("score") for pred in predictions])
+        exp_scores = np.exp(scores - np.max(scores))
+        probs = exp_scores / exp_scores.sum()
+
+        # Include the probabilities in our predictions.
+        for prob, pred in zip(probs, predictions):
+            pred["probability"] = prob
+
+        # Pick the best prediction and set the probability for the null answer.
+        all_predictions[example["id"]] = predictions[0]["text"]
+        if version_2_with_negative:
+            scores_diff_json[example["id"]] = float(min_null_score)
+
+        # Make `predictions` JSON-serializable by casting np.float back to float.
+        all_nbest_json[example["id"]] = [
+            {k: (float(v) if isinstance(v, (np.float32, np.float64)) else v) for k, v in pred.items()}
+            for pred in predictions
+        ]
+
+    # If we have an output_dir, let's save all those dicts.
+    if output_dir is not None:
+        assert os.path.isdir(output_dir), f"{output_dir} is not a directory."
+
+        prediction_file = os.path.join(
+            output_dir, "predictions.json" if prefix is None else f"predictions_{prefix}".json
+        )
+        nbest_file = os.path.join(
+            output_dir, "nbest_predictions.json" if prefix is None else f"nbest_predictions_{prefix}".json
+        )
+        if version_2_with_negative:
+            null_odds_file = os.path.join(
+                output_dir, "null_odds.json" if prefix is None else f"null_odds_{prefix}".json
+            )
+
+        print(f"Saving predictions to {prediction_file}.")
+        with open(prediction_file, "w") as writer:
+            writer.write(json.dumps(all_predictions, indent=4) + "\n")
+        print(f"Saving nbest_preds to {nbest_file}.")
+        with open(nbest_file, "w") as writer:
+            writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+        if version_2_with_negative:
+            print(f"Saving null_odds to {null_odds_file}.")
+            with open(null_odds_file, "w") as writer:
+                writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions, scores_diff_json
@@ -0,0 +1,28 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# Research projects
+
+This folder contains various research projects using 🤗 Transformers. They are not maintained and require a specific
+version of 🤗 Transformers that is indicated in the requirements file of each folder. Updating them to the most recent version of the library will require some work.
+
+To use any of them, just run the command
+```
+pip install -r requirements.txt
+```
+inside the folder of your choice.
+
+If you need help with any of those, contact the author(s), indicated at the top of the `README` of each folder.
@@ -0,0 +1 @@
+transformers == 3.5.1
@@ -0,0 +1 @@
+transformers == 3.5.1
@@ -1,4 +1,4 @@
-transformers
+transformers == 3.5.1

 # For ROUGE
 nltk
@@ -0,0 +1 @@
+transformers == 3.5.1
@@ -0,0 +1 @@
+transformers == 3.5.1
Author	SHA1	Message	Date
Stas Bekman	5ea787a2d8	wip	2020-12-11 14:23:07 -08:00
Stas Bekman	a019424ff0	wip	2020-12-11 14:20:02 -08:00
Stas Bekman	49f7744aa3	wip	2020-12-11 14:17:17 -08:00
Stas Bekman	5455191606	shorten job	2020-12-11 13:53:53 -08:00
Stas Bekman	8507bb7352	set +e	2020-12-11 13:53:12 -08:00
Stas Bekman	f339c16a38	wip	2020-12-11 12:58:01 -08:00
Stas Bekman	40a6227fe1	workflows	2020-12-11 12:56:33 -08:00
Stas Bekman	09cd2464db	workflow restore	2020-12-11 12:55:15 -08:00
Stas Bekman	6a2d0b4e50	test with just 2 jobs	2020-12-11 12:53:54 -08:00
Stas Bekman	eacbd879c9	[wip] [ci] experiment for documentation please ignore	2020-12-11 12:49:31 -08:00
Patrick von Platen	9cc9f4122e	Make ProphetNetModel really compatible with EncoderDecoder (#9033 ) * improve * finish * upload model * fix lm head * fix test	2020-12-11 16:59:54 +01:00
dependabot[bot]	24f6cdeab6	Bump notebook in /examples/research_projects/movement-pruning/lxmert (#9062 ) Bumps [notebook](https://github.com/jupyter/jupyterhub) from 6.1.4 to 6.1.5. - [Release notes](https://github.com/jupyter/jupyterhub/releases) - [Changelog](https://github.com/jupyterhub/jupyterhub/blob/master/CHECKLIST-Release.md) - [Commits](https://github.com/jupyter/jupyterhub/commits) Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>	2020-12-11 10:32:43 -05:00
Lysandre Debut	91fa707217	Remove docs only check (#9065 )	2020-12-11 10:27:31 -05:00
Sylvain Gugger	70527ba694	Fix PreTrainedTokenizer.pad when first inputs are empty (#9018 ) * Fix PreTrainedTokenizer.pad when first inputs are empty * Handle empty inputs case	2020-12-11 10:25:00 -05:00
Sylvain Gugger Lysandre Debut Thomas Wolf	783d7d2629	Reorganize examples (#9010 ) * Reorganize example folder * Continue reorganization * Change requirements for tests * Final cleanup * Finish regroup with tests all passing * Copyright * Requirements and readme * Make a full link for the documentation * Address review comments * Apply suggestions from code review Co-authored-by: Lysandre Debut <lysandre@huggingface.co> * Add symlink * Reorg again * Apply suggestions from code review Co-authored-by: Thomas Wolf <thomwolf@users.noreply.github.com> * Adapt title * Update to new strucutre * Remove test * Update READMEs Co-authored-by: Lysandre Debut <lysandre@huggingface.co> Co-authored-by: Thomas Wolf <thomwolf@users.noreply.github.com>	2020-12-11 10:07:02 -05:00
Suraj Patil	86896de064	update tatoeba workflow (#9051 )	2020-12-11 20:29:15 +05:30
Ganesh KharadandJulien Chaumond	7c8f5f6487	Create README.md (#8096 ) * Create README.md * Fix model card Co-authored-by: Julien Chaumond <julien@huggingface.co>	2020-12-11 09:45:12 -05:00
RamonMamonandJulien Chaumond	5527f78721	Create README.md (#8281 ) * Create README.md * Update model_cards/kiri-ai/distiluse-base-multilingual-cased-et/README.md Co-authored-by: Julien Chaumond <chaumond@gmail.com>	2020-12-11 09:41:29 -05:00
joanginesandJulien Chaumond	c615df7422	Create README.md (#8751 ) * Create README.md * Update model_cards/Cinnamon/electra-small-japanese-generator/README.md Co-authored-by: Julien Chaumond <chaumond@gmail.com>	2020-12-11 09:40:14 -05:00
Ahmed Abdelali Ahmed Abdelali Julien Chaumond	76df559383	QARiB Arabic and dialects models (#8796 ) * Add QARiB models * fix README.md * Fix README.md * Fix README.md * Fix README.md * Fix QARiB files * add models card for QARiB models 860k, 1790k, and 1970k * try to fix PR * re-add files * links aren't allowed here :) Co-authored-by: Ahmed Abdelali <aabdelali@hbku.edu.qa> Co-authored-by: Julien Chaumond <julien@huggingface.co>	2020-12-11 09:38:38 -05:00
moniquebm	b161f1ae54	Update README.md (#8820 )	2020-12-11 09:24:21 -05:00
Panggi Libersa Jasri Akadol	649d389dab	Initial README for `t5-base-indonesian-summarization-cased` model (#9028 ) * Create README.md Initial README for `t5-base-indonesian-summarization-cased` model * Update README for t5-base-indonesian-summarization-cased Typo in README, change from `small` to `base`	2020-12-11 09:18:16 -05:00
Panggi Libersa Jasri Akadol	5e794b6628	Create README.md (#9030 ) Initial README for `t5-small-indonesian-summarization-cased` model	2020-12-11 09:17:29 -05:00
Cola	935e346959	🎨 Change nn.dropout to layer.Dropout (#9047 )	2020-12-11 10:40:25 +01:00
Julien Plu	b01ddc9577	Remove value error (#8985 ) * Remove value error * Try a fix for parameter ordering * Restore previous behavior * Add documentation * Review the comment	2020-12-10 17:17:19 -05:00
NatLun137	91ab02af28	Fix typo #9012 (#1 ) (#9038 ) There is a tiny typo in the code "transformers/examples/language-modeling/run_mlm_wwm.py" at line 284. [Details.](https://github.com/huggingface/transformers/issues/9012)	2020-12-10 16:41:00 -05:00
Sylvain Gugger	8d4bb02056	Refactor FLAX tests (#9034 )	2020-12-10 15:57:39 -05:00
Sylvain Gugger	1310e1a758	Enforce all objects in the main init are documented (#9014 )	2020-12-10 11:57:12 -05:00
Sylvain Gugger	51e81e5895	MPNet copyright files (#9015 )	2020-12-10 09:29:38 -05:00
Sylvain Gugger	35bffd70e2	Fix documention of book in LayoutLM (#9017 )	2020-12-10 09:28:49 -05:00
Cola	c95de29e31	✏️ Fix typo (#9020 )	2020-12-10 08:22:52 +01:00
Stas Bekman	5e637e6c69	[wip] [ci] doc-job-skip take #4 dry-run (#8980 ) * ci-doc-job-skip-take-4 * wip * wip * wip * wip * skip yaml * wip * wip * wip * wip * wip * wip * wip * wip * wip * wip * wip * wip * ready to test * yet another way * trying with HEAD * trying with head.sha * trying with head.sha fix * trying with head.sha fix wip * undo * try to switch to sha * current branch * current branch * PR number check * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride * joy ride	2020-12-09 15:36:36 -05:00
Patrick von Platen	06971ac4f9	[Bart] Refactor - fix issues, consistency with the library, naming (#8900 ) * remove make on the fly linear embedding * start refactor * big first refactor * save intermediate * save intermediat * correct mask issue * save tests * refactor padding masks * make all tests pass * further refactor * make pegasus test pass * fix bool if * fix leftover tests * continue * bart renaming * delete torchscript test hack * fix imports in tests * correct shift * fix docs and repo cons * re-add fix for FSTM * typo in test * fix typo * fix another typo * continue * hot fix 2 for tf * small fixes * refactor types linting * continue * finish refactor * fix import in tests * better bart names * further refactor and add test * delete hack * apply sylvains and lysandres commens * small perf improv * further perf improv * improv perf * fix typo * make style * small perf improv	2020-12-09 20:55:24 +01:00
Funtowicz MorganandMarc van Zee	75627148ee	Flax Masked Language Modeling training example (#8728 ) * Remove "Model" suffix from Flax models to look more 🤗 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Initial working (forward + backward) for Flax MLM training example. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Simply code Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Addressing comments, using module and moving to LM task. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Restore parameter name "module" wrongly renamed model. Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Restore correct output ordering... Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Actually commit the example 😅 Signed-off-by: Morgan Funtowicz <morgan@huggingface.co> * Add FlaxBertModelForMaskedLM after rebasing. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Make it possible to initialize the training from scratch Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Reuse flax linen example of cross entropy loss Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Added specific data collator for flax Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Remove todo for data collator Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Added evaluation step Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Added ability to provide dtype to support bfloat16 on TPU Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Enable flax tensorboard output Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Enable jax.pmap support. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Ensure batches are correctly sized to be dispatched with jax.pmap Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Enable bfloat16 with --fp16 cmdline args Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Correctly export metrics to tensorboard Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Added dropout and ability to use it. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Effectively enable & disable during training and evaluation steps. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Oops. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Enable specifying kernel initializer scale Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Style. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Added warmup step to the learning rate scheduler. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Fix typo. Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Print training loss Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Make style Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * fix linter issue (flake8) Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Fix model matching Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Fix dummies Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Fix non default dtype on Flax models Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Use the same create_position_ids_from_input_ids for FlaxRoberta Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Make Roberta attention as Bert Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * fix copy Signed-off-by: Morgan Funtowicz <funtowiczmo@gmail.com> * Wording. Co-authored-by: Marc van Zee <marcvanzee@gmail.com> Co-authored-by: Marc van Zee <marcvanzee@gmail.com>	2020-12-09 17:13:56 +01:00
StillKeepTry	df2af6d8b8	Add MP Net 2 (#9004 )	2020-12-09 10:32:43 -05:00
cronoik	8729109855	fixes #8968 (#9009 )	2020-12-09 16:21:41 +01:00
Simon Brandeis	e977ed2142	Add the code_search_net dataset tag to CodeBERTa model cards (#9005 )	2020-12-09 15:43:19 +01:00
Patrick von Platen	da37a21c89	push (#9008 )	2020-12-09 15:14:33 +01:00
Sylvain Gugger	61abd50b98	Remove use of deprected method in Trainer HP search (#8996 )	2020-12-09 09:13:41 -05:00
Sylvain Gugger	7e1d709e2a	Fix link to stable version in the doc navbar (#9007 )	2020-12-09 09:11:39 -05:00
Patrick von Platen Ayush Jain ayushtiku5	02d0e0355c	Diverse beam search 2 (#9006 ) * diverse beam search * bug fixes * bug fixes * bug fix * separate out diverse_beam_search function * separate out diverse_beam_search function * bug fix * improve code quality * bug fix * bug fix * separate out diverse beam search scorer * code format * code format * code format * code format * add test * code format * documentation changes * code quality * add slow integration tests * more general name * refactor into logits processor * add test * avoid too much copy paste * refactor * add to docs * fix-copies * bug fix * Revert "bug fix" This reverts commit c99eb5a8dc57a7b0d33a8ac06d8c6a32a7812ad4. * improve comment * implement sylvains feedback Co-authored-by: Ayush Jain <a.jain@sprinklr.com> Co-authored-by: ayushtiku5 <40797286+ayushtiku5@users.noreply.github.com>	2020-12-09 15:00:37 +01:00
Lysandre Debut	67ff1c314a	Templates overhaul 1 (#8993 )	2020-12-08 18:00:07 -05:00
Sylvain GuggerandPatrick von Platen	447808c85f	New squad example (#8992 ) * Add new SQUAD example * Same with a task-specific Trainer * Address review comment. * Small fixes * Initial work for XLNet * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Final clean up and working XLNet script * Test and debug * Final working version * Add new SQUAD example * Same with a task-specific Trainer * Address review comment. * Small fixes * Initial work for XLNet * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Final clean up and working XLNet script * Test and debug * Final working version * Add tick * Update README * Address review comments Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>	2020-12-08 14:39:29 -05:00
guillaume-be	7809eb82ae	Removed unused `encoder_hidden_states` and `encoder_attention_mask` (#8972 ) * Removed unused `encoder_hidden_states` and `encoder_attention_mask` from MobileBert * Removed decoder tests for MobileBert * Removed now unnecessary import	2020-12-08 12:04:34 -05:00
Lysandre Debut	b7cdd00f15	Fix interaction of return_token_type_ids and add_special_tokens (#8854 )	2020-12-08 12:04:01 -05:00
Sylvain Gugger	04c446f764	Make `ModelOutput` pickle-able (#8989 )	2020-12-08 11:59:40 -05:00
Julien Chaumond	0d9e6ca9ed	[model_card] remove bogus testing changes	2020-12-08 09:58:45 -05:00