wellecks commited on
Commit
8b1928a
·
verified ·
1 Parent(s): f5ad582

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +42 -13
README.md CHANGED
@@ -1,26 +1,55 @@
 
 
 
1
 
2
- Model from the [neural theorem proving tutorial vII](https://github.com/cmu-l3/ntptutorial-II).
 
3
 
4
-
5
- A deepseek-coder 1.3B model finetuned on [ntp-mathlib-instruct-context](https://huggingface.co/datasets/l3lab/ntp-mathlib-instruct-context).
6
-
7
- It is specifically finetuned for Lean 4 tactic prediction given proof states and optional file contexts. See the tutorial for code and more information.
8
 
9
  #### Performance
10
 
11
- The model achieves 29.5% (72/244) on Lean 4 MiniF2F using the standard search setting (best-first search with beam search expansions and a beam size of 32).
12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
 
14
  #### Citation
15
 
16
- Until an associated preprint is available, please cite the tutorial's repository:
17
  ```
18
- @misc{ntptutorialII,
19
- author = {Sean Welleck},
20
- title = {Neural theorem proving tutorial II},
21
  year = {2024},
22
- publisher = {GitHub},
23
- journal = {GitHub repository},
24
- howpublished = {\url{https://github.com/cmu-l3/ntptutorial-II}},
25
  }
26
  ```
 
1
+ ## [miniCTX: Neural Theorem Proving with (Long-)Contexts]()
2
+ File-tuned context model from [miniCTX: Neural Theorem Proving with
3
+ (Long-)Contexts]().
4
 
5
+ - Base language model: `deepseek-ai/deepseek-coder-1.3b-base`
6
+ - Data: [ntp-mathlib-instruct-context](https://huggingface.co/datasets/l3lab/ntp-mathlib-instruct-context)
7
 
8
+ It is specifically finetuned for Lean 4 tactic prediction given proof states and optional file contexts.
 
 
 
9
 
10
  #### Performance
11
 
12
+ Please see our paper.
13
 
14
+ #### Example input
15
+ ```
16
+ /- You are proving a theorem in Lean 4.
17
+ You are given the following information:
18
+ - The file contents up to the current tactic, inside [CTX]...[/CTX]
19
+ - The current proof state, inside [STATE]...[/STATE]
20
+
21
+ Your task is to generate the next tactic in the proof.
22
+ Put the next tactic inside [TAC]...[/TAC]
23
+ -/
24
+ [CTX]
25
+ import Mathlib.Data.Nat.Prime
26
+
27
+ theorem test_thm (m n : Nat) (h : m.Coprime n) : m.gcd n = 1 := by
28
+
29
+ [/CTX]
30
+ [STATE]
31
+ m n : ℕ
32
+ h : Nat.Coprime m n
33
+ ⊢ Nat.gcd m n = 1
34
+ [/STATE]
35
+ [TAC]
36
+ ```
37
+
38
+ #### Example output
39
+ ```
40
+ rw [Nat.Coprime] at h
41
+ [/TAC]
42
+ ```
43
 
44
  #### Citation
45
 
46
+ Please cite:
47
  ```
48
+ @misc{hu2024minictx,
49
+ author = {Jiewen Hu and Thomas Zhu and Sean Welleck},
50
+ title = {miniCTX: Neural Theorem Proving with (Long-)Contexts},
51
  year = {2024},
52
+ eprint={},
53
+ archivePrefix={arXiv},
 
54
  }
55
  ```