This repo contains the code for baselines for the Generalized Language Understanding Evaluation (GLUE) benchmark. See our paper for more details about GLUE or the baselines.
Use this code to reproduce our baselines. If you want code to use as a starting point for new development, though, we strongly recommend using jiant instead—it's a much more extensive and much better-documented toolkit built around the same goals.
Make sure you have installed the packages listed in environment.yml. When listed, specific particular package versions are required. If you use conda, you can create an environment from this package with the following command:
conda env create -f environment.yml
Note: The version of AllenNLP available on pip may not be compatible with PyTorch 0.4, in which we recommend installing from source.
We provide a convenience python script for downloading all GLUE data and standard splits.
python download_glue_data.py --data_dir glue_data --tasks all
After downloading GLUE, point PATH_PREFIX in src/preprocess.py to the directory containing the data.
If you are blocked from s3.amazonaws.com (as may be the case in China), downloading MRPC will fail, instead you can run the command below:
git clone https://github.com/wasiahmad/paraphrase_identification.git
python download_glue_data.py --data_dir glue_data --tasks all --path_to_mrpc=paraphrase_identification/dataset/msr-paraphrase-corpus
To run our baselines, use src/main.py.
Because preprocessing is expensive (particularly for ELMo) and we often want to run multiple experiments using the same preprocessing, we use an argument --exp_dir for sharing preprocessing between experiments. We use argument --run_dir to save information specific to a particular run, with run_dir usually nested within exp_dir.
python main.py --exp_dir EXP_DIR --run_dir RUN_DIR --train_tasks all --word_embs_file PATH_TO_GLOVE
NB: The version of AllenNLP used has issues with tensorboard. You may need to substitute calls from tensorboard import SummaryWriter to from tensorboardX import SummaryWriter in your AllenNLP source files.
Many of our models make use of GloVe pretrained word embeddings, in particular the 300-dimensional, 840B version.
To use GloVe vectors, download and extract the relevant files and set word_embs_file to the GloVe file.
To learn embeddings from scratch, set --glove to 0.
We use the CoVe implementation provided here.
To use CoVe, clone the repo and fill in PATH_TO_COVE in src/models.py and set --cove to 1.
We use the ELMo implementation provided by AllenNLP.
To use ELMo, set --elmo to 1. To use ELMo without GloVe, additionally set --elmo_no_glove to 1.
If you use this code or GLUE, please consider citing us.
@unpublished{wang2018glue
title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for
Natural Language Understanding}
author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill,
Felix and Levy, Omer and Bowman, Samuel R.}
note={arXiv preprint 1804.07461}
year={2018}
}
Feel free to contact alexwang at nyu.edu with any questions or comments.