/coproof

Primary LanguageJupyter NotebookGNU General Public License v2.0GPL-2.0

CoProver

CoProver is a proof recommendation system for proof assistants trained on PVS and Imandra

Installation

CoProver is intended to run on Python 3.9+. Once an environment has been selected, install the dependencies via setup.py

cd coproof
pip install -e .

Tasks

Training and running the PVS Command Predictor experiments

To train and evaluate the PVS command prediction, execute this module,

python -m coprover.cmdpred.train_t5

Note this uses PyTorch Lightning to do DDP training across several GPUs. You may need to adjust the num_gpus argument to match the number of GPUs you have on your system, or use 0 for software only.

The baselines for comparison can be executed via,

python -m coprover.cmdpred.run_baselines

Training and running the PVS Lemma Retrieval experiments

To first train the underlying Sentence BERT model,

python -m coprover.lemmaret.bin.train_sbert

Generating JSON formatted data from PVS JSON proof traces

python -m coprover.feats.featurize_cmdpred