Awesome

WinoGrande

Version 1.1

Data

Download dataset by download_winogrande.sh

./data/
├── train_[xs,s,m,l,xl].jsonl          # training set with differnt sizes
├── train_[xs,s,m,l,xl]-labels.lst     # answer labels for training sets
├── dev.jsonl                          # development set
├── dev-labels.lst                     # answer labels for development set
├── test.jsonl                         # test set
├── sample-submissions-labels.lst      # example submission file for leaderboard    
└── eval.py                            # evaluation script

You can use train_*.jsonl for training models and dev for validation. Please note that labels are not included in test.jsonl. To evaluate your models on test set, make a submission to our leaderboard.

Run experiments

Setup

1. Download dataset by download_winogrande.sh
2. pip install -r requirements.txt

Training (fine-tuning)

1. You can train your model by ./scripts/run_experiment.py (see sample_training.sh).

    e.g., 
    export PYTHONPATH=$PYTHONPATH:$(pwd)
   
    python scripts/run_experiment.py \
    --model_type roberta_mc \ 
    --model_name_or_path roberta-large \
    --task_name winogrande \
    --do_eval \
    --do_lower_case \
    --data_dir ./data \
    --max_seq_length 80 \
    --per_gpu_eval_batch_size 4 \
    --per_gpu_train_batch_size 16 \
    --learning_rate 1e-5 \
    --num_train_epochs 3 \
    --output_dir ./output/models/ \
    --do_train \
    --logging_steps 4752 \
    --save_steps 4750 \
    --seed 42 \
    --data_cache_dir ./output/cache/ \
    --warmup_pct 0.1 \
    --evaluate_during_training
   

2. If you have an access to beaker, you can run your experiments by sh ./train_winogrande_on_bkr.sh.

3. Results will be stored under ./output/models/.

Prediction (on the test set)

1. You can make predictions by ./scripts/run_experiment.py directly (see sample_prediction.sh).

    e.g., 
    export PYTHONPATH=$PYTHONPATH:$(pwd)
   
    python scripts/run_experiment.py \
    --model_type roberta_mc \
    --model_name_or_path .output/models \
    --task_name winogrande \
    --do_predict \
    --do_lower_case \
    --data_dir ./data \
    --max_seq_length 80 \
    --per_gpu_eval_batch_size 4 \
    --output_dir ./output/models/ \
    --data_cache_dir ./output/cache/ \
   

2. If you have an access to beaker, you can run your experiments by sh ./predict_winogrande_on_bkr.sh.

3. Result is stored in ./output/models/predictions_test.lst

Evaluation

You can use eval.py for evaluation on the dev split, which yields metrics.json.

e.g., python eval.py --preds_file ./YOUR_PREDICTIONS.lst --labels_file ./dev-labels.lst

In the prediction file, each line consists of the predictions (1 or 2) by 5 training sets (ordered by xs, s, m, l, xl, separated by comma) for each evauation set question.

 2,1,1,1,1
 1,1,2,2,2
 1,1,1,1,1
 .........
 .........

Namely, the first column is the predictions by a model trained/finetuned on train_xs.jsonl, followed by a model prediction by train_s.jsonl, ... , and the last (fifth) column is the predictions by a model from train_xl.jsonl. Please checkout a sample submission file (sample-submission-labels.lst) for reference.

Submission to Leaderboard

You can submit your predictions on test set to the leaderboard. The submission file must be named as predictions.lst. The format is the same as above.

Reference

If you use this dataset, please cite the following paper:

@article{sakaguchi2019winogrande,
    title={WinoGrande: An Adversarial Winograd Schema Challenge at Scale},
    author={Sakaguchi, Keisuke and Bras, Ronan Le and Bhagavatula, Chandra and Choi, Yejin},
    journal={arXiv preprint arXiv:1907.10641},
    year={2019}
}

License

Winogrande (codebase) is licensed under the Apache License 2.0. The dataset is licensed under CC-BY.

Questions?

Please file GitHub issues with your questions/suggestions. You may also ask us questions at our google group.

Contact

Email: keisukes[at]allenai.org