README.md

llemma formal2formal

Scripts for the Lean formal2formal (tactic prediction) experiments in
Llemma: an open language model for mathematics [Azerbayev et al 2023]

Setup

Install Python packages:

pip install -r requirements.txt

Install Lean:

# from https://leanprover-community.github.io/install/linux.html
# After running this command, select (2), then `nightly`, then `y`:
curl https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh -sSf | sh
source $HOME/.elan/env
lake

Configure LeanDojo:

export CONTAINER="native"

Run

See scripts

Compute metrics

python compute_metrics.py
==>

codellama7b_minif2f_test        0.20491803278688525     50      244
codellama34b_minif2f_test       0.22131147540983606     54      244
llemma7b_minif2f_test   0.26229508196721313     64      244
llemma34b_minif2f_test  0.2581967213114754      63      244

Troubleshooting

• We observe a Ray error when running the 34b script (with VLLM --tp-degree > 1) on an untraced LeanDojo repo. A workaround is to run the 7b script with --tp-degree 1 such that LeanDojo completes tracing the repo. Then run the 34b script with --tp-degree > 1.

Citation

Please cite the following:

@article{azerbayev2023llemma,
    title={Llemma: an open language model for mathematics},
    author={Zhangir Azerbayev and Hailey Schoelkopf and Keiran Paster and Marco Dos Santos and Stephen McAleer and Albert Q. Jiang and Jia Deng and Stella Biderman and Sean Welleck},
    eprint={xyz.xyz},
    archivePrefix={arXiv}
    year={2023}
}