MAUD Supplementary Extraction Task

This repository contains the dataset and baseline code for the MAUD supplementary extraction task, as described in the appendix of "MAUD: An Expert-Annotated Legal NLP Dataset for Merger Agreement Understanding".

For the main MAUD dataset and baselines, see github.com/TheAtticusProject/maud.

bugs: The baselines reported in the papers are underperforming due to a training bug. See #1 .

Installation

pip install torch transformers tensorboard pandas scikit-learn tqdm

Notes

During the first run, feature caching and evaluation requires a lot of CPU memory (>=150 GB) and will save about 25 GB of files on the hard disk. This CPU requirement can be reduced, at the expense of speed, by lowering the --threads count in run_maud.sh.

Training uses around 22 GB of GPU memory.

Validation runs (grid-search)

./run_maud.sh

Test runs with best-performing hyperparameters

./run_maud_best_hp.sh