Model from the neural theorem proving tutorial vII.
A deepseek-coder 1.3B model finetuned on ntp-mathlib-instruct-st for 3 epochs.
It is specifically finetuned for Lean 4 tactic prediction given proof states. See the tutorial for code and more information.
Performance
The model achieves 29.1% (71/244) on Lean 4 MiniF2F using the standard search setting (best-first search with beam search expansions and a beam size of 32).
Citation
Until an associated preprint is available, please cite the tutorial's repository:
@misc{ntptutorialII,
author = {Sean Welleck},
title = {Neural theorem proving tutorial II},
year = {2024},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/cmu-l3/ntptutorial-II}},
}