aashish1904
commited on
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
---
|
3 |
+
|
4 |
+
pipeline_tag: text-generation
|
5 |
+
license: other
|
6 |
+
language:
|
7 |
+
- en
|
8 |
+
tags:
|
9 |
+
- math
|
10 |
+
|
11 |
+
---
|
12 |
+
|
13 |
+
![](https://lh7-rt.googleusercontent.com/docsz/AD_4nXeiuCm7c8lEwEJuRey9kiVZsRn2W-b4pWlu3-X534V3YmVuVc2ZL-NXg2RkzSOOS2JXGHutDuyyNAUtdJI65jGTo8jT9Y99tMi4H4MqL44Uc5QKG77B0d6-JfIkZHFaUA71-RtjyYZWVIhqsNZcx8-OMaA?key=xt3VSDoCbmTY7o-cwwOFwQ)
|
14 |
+
|
15 |
+
# QuantFactory/internlm2-step-prover-GGUF
|
16 |
+
This is quantized version of [internlm/internlm2-step-prover](https://huggingface.co/internlm/internlm2-step-prover) created using llama.cpp
|
17 |
+
|
18 |
+
# Original Model Card
|
19 |
+
|
20 |
+
|
21 |
+
# InternLM-Step-Prover
|
22 |
+
|
23 |
+
<div align="center">
|
24 |
+
|
25 |
+
<img src="https://raw.githubusercontent.com/InternLM/InternLM/main/assets/logo.svg" width="200"/>
|
26 |
+
<div> </div>
|
27 |
+
<div align="center">
|
28 |
+
<b><font size="5">InternLM-Math</font></b>
|
29 |
+
<sup>
|
30 |
+
<a href="https://internlm.intern-ai.org.cn/">
|
31 |
+
<i><font size="4">HOT</font></i>
|
32 |
+
</a>
|
33 |
+
</sup>
|
34 |
+
<div> </div>
|
35 |
+
</div>
|
36 |
+
|
37 |
+
A state-of-the-art LEAN4 step prover.
|
38 |
+
|
39 |
+
[💻 Github](https://github.com/InternLM/InternLM-Math) [📊Dataset](https://huggingface.co/datasets/internlm/Lean-Github) [📖 Paper](https://arxiv.org/abs/2407.17227)
|
40 |
+
</div>
|
41 |
+
|
42 |
+
InternLM-Step-Prover is a 7B language model trained on Lean-Github and multiple sythesis datasets. InternLM-Step-Prover achieves state-of-the-art performances on MiniF2F, ProofNet, and Putnam math benchmarks, showing its formal math proving ability in multiple domains.
|
43 |
+
|
44 |
+
# Dialogue Example
|
45 |
+
```
|
46 |
+
### Input Example
|
47 |
+
DECL MyNat.mul_pow
|
48 |
+
GOAL a b n : N
|
49 |
+
⊢ (a * b) ^ n = a ^ n * b ^ n
|
50 |
+
### Output Example
|
51 |
+
PROOFSTEP induction n with t Ht
|
52 |
+
```
|
53 |
+
|
54 |
+
# Performance
|
55 |
+
|
56 |
+
## MiniF2F
|
57 |
+
| Method | Model size | Pass | miniF2F-valid | miniF2F-test |
|
58 |
+
|--------|------------|------|---------------|--------------|
|
59 |
+
| **Whole-Proof Generation Methods** |
|
60 |
+
| GPT-4-turbo 0409 | - | 64 | 25.4% | 23.0% |
|
61 |
+
| DeepSeekMath-Base | 7B | 128 | 25.4% | 27.5% |
|
62 |
+
| DeepSeek-Prover | 7B | 1 | - | 30.0% |
|
63 |
+
| | | 64 | - | 46.3% |
|
64 |
+
| | | 128 | - | 46.3% |
|
65 |
+
| | | 8192 | - | 48.8% |
|
66 |
+
| | | 65536 | - | 50.0% |
|
67 |
+
| | | cumulative | *60.2%* | *52.0%* |
|
68 |
+
| TheoremLlama | - | cumulative | 36.5% | 33.6% |
|
69 |
+
| **Tree Search Methods** |
|
70 |
+
| COPRA (GPT-3.5) | - | 1 | - | 9.0% |
|
71 |
+
| COPRA (GPT-4) | - | 1 | - | 26.6% |
|
72 |
+
| DSP(Isabelle) | 540B | 100 | 42.6% | 38.9% |
|
73 |
+
| Proof Artifact Co-Training | 837M | 1 | 23.9% | 24.6% |
|
74 |
+
| | | 8 | 29.3% | 29.2% |
|
75 |
+
| ReProver | 229M | 1 | - | 25.0% |
|
76 |
+
| Llemma | 7B | 1 | 26.2% | 26.2% |
|
77 |
+
| Llemma | 34B | 1 | 27.9% | 25.8% |
|
78 |
+
| Curriculum Learning | 837M | 1 | 33.6% | 29.6% |
|
79 |
+
| | | 8 | 41.2% | 34.5% |
|
80 |
+
| | | 64 | 47.3% | 36.6% |
|
81 |
+
| Hypertree Proof Search | 600M | cumulative | 58.6% | - |
|
82 |
+
| | | 64 | - | 41.0% |
|
83 |
+
| Lean-STaR | 7B | 64 | - | 46.3% |
|
84 |
+
| InternLM2-Math | 7B | 1 | 29.9% | 30.3% |
|
85 |
+
| InternLM2-Math-Plus | 7B | 1 | - | 43.4% |
|
86 |
+
| InternLM2-Step-Prover | 7B | 1 | 59.8% | 48.8% |
|
87 |
+
| InternLM2-Step-Prover | 7B | 64 | **63.9%** | **54.5%** |
|
88 |
+
|
89 |
+
## Proofnet & Putnam
|
90 |
+
| Method | Model size | Pass | result |
|
91 |
+
|--------|------------|------|--------|
|
92 |
+
| **ProofNet benchmark** |
|
93 |
+
| ReProver | 229M | 1 | 13.8% |
|
94 |
+
| InternLM2-Step-Prover | 7B | 1 | **18.1%** |
|
95 |
+
| **Putnam benchmark** |
|
96 |
+
| GPT-4 | - | 10 | 1/640 |
|
97 |
+
| COPRA (GPT-4) | - | 10 | 1/640 |
|
98 |
+
| DSP(Isabelle) | 540B | 10 | 4/640 |
|
99 |
+
| ReProver | 229M | 1 | 0/640 |
|
100 |
+
| InternLM2-Step-Prover | 7B | 1 | **5/640** |
|
101 |
+
|
102 |
+
# Citation and Tech Report
|
103 |
+
```
|
104 |
+
@misc{wu2024leangithubcompilinggithublean,
|
105 |
+
title={LEAN-GitHub: Compiling GitHub LEAN repositories for a versatile LEAN prover},
|
106 |
+
author={Zijian Wu and Jiayu Wang and Dahua Lin and Kai Chen},
|
107 |
+
year={2024},
|
108 |
+
eprint={2407.17227},
|
109 |
+
archivePrefix={arXiv},
|
110 |
+
primaryClass={cs.AI},
|
111 |
+
url={https://arxiv.org/abs/2407.17227},
|
112 |
+
}
|
113 |
+
```
|