README.md

metadata

tags:
  - theorem-proving
  - math
  - lean

miniCTX: Neural Theorem Proving with (Long-)Contexts

Lean 4 tactic prediction examples extracted from Mathlib.

These examples have not been formatted for instruction tuning (including data splits).

Please see l3lab/ntp-mathlib-instruct-* for datasets with instruction tuning examples.

Version

Generated using ntptoolkit's ntp-training-data.

It used the following config for ntp-training-data:

{
    "repo": "https://github.com/leanprover-community/mathlib4",
    "commit": "cf8e23a62939ed7cc530fbb68e83539730f32f86",
    "lean": "leanprover/lean4:v4.4.0",
    "name": "mathlib",
    "import_file": "Mathlib.lean",
    "imports": ["Mathlib"]
}

Example usage:

ds = datasets.load_dataset('l3lab/ntp-mathlib')

print(len(ds['train']))
# ==> 307049

Format:

{
  'state': 'proof state',
  'srcUpToTactic': 'source up to tactic invocation',
  'nextTactic': 'tactic',
  'declUpToTactic': 'declariation up to tactic invocation',
  'declId': 'unique ID for declaration',
  'decl': 'declaration',
  'file_tag': 'file ID'
}

Citation

Please cite:

@misc{hu2024minictx,
  title={miniCTX: Neural Theorem Proving with (Long-)Contexts}, 
  author={Jiewen Hu and Thomas Zhu and Sean Welleck},
  year={2024},
  eprint={2408.03350},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2408.03350}, 
}