metadata
pipeline_tag: text-generation
license: other
language:
- en
tags:
- math
datasets:
- internlm/Lean-Workbook
- internlm/Lean-Github
InternLM2.5-Step-Prover
InternLM2.5-Step-Prover-Critic is a 1.8B critic model which achieves state-of-the-art performances on MiniF2F, ProofNet, and Putnam math benchmarks, showing its formal math proving ability in multiple domains.
Dialogue Example
import torch
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained(
"internlm/internlm2_5-step-prover-critic",
device_map="cuda",
torch_dtype=torch.float16,
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("internlm/internlm2_5-step-prover-critic", trust_remote_code=True)
chat_1 = [
{"role": "user", "content": "Which state is closer to 'no goals'?"},
{"role": "assistant", "content": "no goals"}
]
chat_2 = [
{"role": "user", "content": "Which state is closer to 'no goals'?"},
{"role": "assistant", "content": "x : ℕ\nh₀ : ↑x + 4 / 100 * ↑x = 598\n⊢ 100 * x = 100 * 575"}
]
score1 = model.get_score(tokenizer, chat_1)
score2 = model.get_score(tokenizer, chat_2)
print("score1: ", score1)
print("score2: ", score2)
Performance
MiniF2F
| Method | Model size | Pass | miniF2F-valid | miniF2F-test |
|---|---|---|---|---|
| Whole-Proof Generation Methods | ||||
| GPT-4-turbo 0409 | - | 64 | 25.4% | 23.0% |
| DeepSeekMath-Base | 7B | 128 | 25.4% | 27.5% |
| DeepSeek-Prover | 7B | 1 | - | 30.0% |
| 64 | - | 46.3% | ||
| 128 | - | 46.3% | ||
| 8192 | - | 48.8% | ||
| 65536 | - | 50.0% | ||
| cumulative | 60.2% | 52.0% | ||
| DeepSeek-Prover-1.5 | 7B | 32 | - | 63.5% |
| TheoremLlama | - | cumulative | 36.5% | 33.6% |
| Tree Search Methods | ||||
| COPRA (GPT-3.5) | - | 1 | - | 9.0% |
| COPRA (GPT-4) | - | 1 | - | 26.6% |
| DSP(Isabelle) | 540B | 100 | 42.6% | 38.9% |
| Proof Artifact Co-Training | 837M | 1 | 23.9% | 24.6% |
| 8 | 29.3% | 29.2% | ||
| ReProver | 229M | 1 | - | 25.0% |
| Llemma | 7B | 1 | 26.2% | 26.2% |
| Llemma | 34B | 1 | 27.9% | 25.8% |
| Curriculum Learning | 837M | 1 | 33.6% | 29.6% |
| 8 | 41.2% | 34.5% | ||
| 64 | 47.3% | 36.6% | ||
| Hypertree Proof Search | 600M | cumulative | 58.6% | - |
| 64 | - | 41.0% | ||
| Lean-STaR | 7B | 64 | - | 46.3% |
| InternLM2-Math | 7B | 1 | 29.9% | 30.3% |
| InternLM2-Math-Plus | 7B | 1 | - | 43.4% |
| InternLM2-Step-Prover | 7B | 1 | 59.8% | 48.8% |
| InternLM2.5-Step-Prover | 7B | 1 | 55.4% | 47.3% |
| InternLM2.5-Step-Prover+Critic | 7B | 256 | 69.6% | 65.9% |
Proofnet & Putnam
| Method | Model size | Pass | result |
|---|---|---|---|
| ProofNet benchmark | |||
| ReProver | 229M | 1 | 13.8% |
| InternLM2-Step-Prover | 7B | 1 | 18.1% |
| InternLM2.5-Step-Prover | 7B | 256 | 27.0% |
| Putnam benchmark | |||
| GPT-4 | - | 10 | 1/640 |
| COPRA (GPT-4) | - | 10 | 1/640 |
| DSP(Isabelle) | 540B | 10 | 4/640 |
| ReProver | 229M | 1 | 0/640 |
| InternLM2-Step-Prover | 7B | 1 | 5/640 |
| InternLM2.5-Step-Prover | 7B | 1 | 6/640 |
Citation and Tech Report
@misc{wu2024internlm25stepproveradvancingautomatedtheorem,
title={InternLM2.5-StepProver: Advancing Automated Theorem Proving via Expert Iteration on Large-Scale LEAN Problems},
author={Zijian Wu and Suozhi Huang and Zhejian Zhou and Huaiyuan Ying and Jiayu Wang and Dahua Lin and Kai Chen},
year={2024},
eprint={2410.15700},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2410.15700},
}