Dataset Cartography

Code for the paper Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics at EMNLP 2020.

This repository contains implementation of data maps, as well as other data selection baselines, along with notebooks for data map visualizations.

If using, please cite:

@inproceedings{swayamdipta2020dataset,
    title={Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics},
    author={Swabha Swayamdipta and Roy Schwartz and Nicholas Lourie and Yizhong Wang and Hannaneh Hajishirzi and Noah A. Smith and Yejin Choi},
    booktitle={Proceedings of EMNLP},
    url={https://arxiv.org/abs/2009.10795},
    year={2020}
}

This repository can be used to build Data Maps, like this one for SNLI using a RoBERTa-Large classifier.

Pre-requisites

This repository is based on the HuggingFace Transformers library.

Train GLUE-style model and compute training dynamics

To train a GLUE-style model using this repository:

python -m cartography.classification.run_glue \
    -c configs/$TASK.jsonnet \
    --do_train \
    --do_eval \
    -o $MODEL_OUTPUT_DIR

The best configurations for our experiments for each of the $TASKs (SNLI, MNLI, QNLI or WINOGRANDE) are provided under configs.

This produces a training dynamics directory $MODEL_OUTPUT_DIR/training_dynamics, see a sample here.

Note: you can use any other set up to train your model (independent of this repository) as long as you produce the dynamics_epoch_$X.jsonl for plotting data maps, and filtering different regions of the data. The .jsonl file must contain the following fields for every training instance:

• guid : instance ID matching that in the original data file, for filtering,
• logits_epoch_$X : logits for the training instance under epoch $X,
• gold : index of the gold label, must match the logits array.

Plot Data Maps

To plot data maps for a trained $MODEL (e.g. RoBERTa-Large) on a given $TASK (e.g. SNLI, MNLI, QNLI or WINOGRANDE):

python -m cartography.selection.train_dy_filtering \
    --plot \
    --task_name $TASK \
    --model_dir $PATH_TO_MODEL_OUTPUT_DIR_WITH_TRAINING_DYNAMICS \
    --model $MODEL_NAME

Data Map Coordinates

The coordinates for producing RoBERTa-Large data maps for SNLI, QNLI, MNLI and WINOGRANDE, as reported in the paper can be found under data/data_map_coordinates/. Each .jsonl file contains the following fields for each instance in the train set:

• guid : instance ID matching that in the original data file,
• index,
• confidence,
• variability,
• correctness.

Data Selection

To select (different amounts of) data based on various metrics from training dynamics:

python -m cartography.selection.train_dy_filtering \
    --filter \
    --task_name $TASK \
    --model_dir $PATH_TO_MODEL_OUTPUT_DIR_WITH_TRAINING_DYNAMICS \
    --metric $METRIC \
    --data_dir $PATH_TO_GLUE_DIR_WITH_ORIGINAL_DATA_IN_TSV_FORMAT

Supported $TASKs include SNLI, QNLI, MNLI and WINOGRANDE, and $METRICs include confidence, variability, correctness, forgetfulness and threshold_closeness; see paper for more details.

To select hard-to-learn instances, set $METRIC as "confidence" and for ambiguous, set $METRIC as "variability". For easy-to-learn instances: set $METRIC as "confidence" and use the flag --worst.

Contact and Reference

For questions and usage issues, please contact [email protected]. If you use dataset cartography for research, please cite our paper as follows:

@inproceedings{swayamdipta-etal-2020-dataset,
    title = "Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics",
    author = "Swayamdipta, Swabha  and
      Schwartz, Roy  and
      Lourie, Nicholas  and
      Wang, Yizhong  and
      Hajishirzi, Hannaneh  and
      Smith, Noah A.  and
      Choi, Yejin",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.746",
    doi = "10.18653/v1/2020.emnlp-main.746",
    pages = "9275--9293",
}

Copyright [2020] [Swabha Swayamdipta]