File size: 2,865 Bytes
11f4234
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f170185
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
---
license: apache-2.0
language:
- en
tags:
- lean
- lean4
- theorem-proving
- mathlib
configs:
- config_name: sft-90
  data_files:
  - split: train
    path: sft-90/train.jsonl
  - split: validation
    path: sft-90/validation.jsonl
  - split: holdout
    path: sft-90/holdout.jsonl
- config_name: sft-1000
  data_files:
  - split: train
    path: sft-1000/train.jsonl
  - split: validation
    path: sft-1000/validation.jsonl
  - split: holdout
    path: sft-1000/holdout.jsonl
- config_name: dpo-2000
  data_files:
  - split: train
    path: dpo-2000/train.jsonl
  - split: validation
    path: dpo-2000/validation.jsonl
  - split: holdout
    path: dpo-2000/holdout.jsonl
- config_name: repair-2800
  data_files:
  - split: train
    path: repair-2800/train.jsonl
  - split: validation
    path: repair-2800/validation.jsonl
- config_name: strict-evaluation-22
  data_files:
  - split: test
    path: strict-evaluation-22/test.jsonl
---

# Kakeya OProver Training

Apache-2.0 training and evaluation assets for the four-stage Kakeya OProver LoRA journey.

## Configs and split counts

- `sft-90`: 63 train / 9 validation / 18 holdout
- `sft-1000`: 700 train / 100 validation / 200 holdout
- `dpo-2000`: 1,400 train / 200 validation / 400 holdout preference pairs
- `repair-2800`: 2,800 train / 400 validation (train: 2,100 repair + 700 rehearsal; validation: 300 + 100)
- `strict-evaluation-22`: 22 public-safe task identifiers and split metadata only (18 holdout + 4 canary); aggregate outcomes are in `strict-evaluation-22/results.json`

## Schema

SFT records include the exact public Lean context, target, and verified proof body. DPO records include `prompt`, `chosen`, `rejected`, rejection metadata, and source bindings. Repair records include the public context, failed proof, structured compiler diagnostic, and corrected proof. Strict evaluation records intentionally omit prompts and candidate proof bodies.

## Provenance and compile validation

All proof-bearing samples are pure Mathlib material or derived negatives/repairs bound to Mathlib revision `360da6fa66c1273b76b6b2d8c5666fd5ac2e3b56` and Lean `v4.32.0-rc1`. Source manifests report compile validation and zero cross-split family overlap where applicable. Provenance-incomplete OProofs mixed samples were excluded.

## Leakage and release policy

Holdout splits were excluded from training. The strict 22-task release contains metadata and aggregates only. No production Jensen prompt, raw candidate proof body, private chain-of-thought, credential, or local path is included.

## License and attribution

Apache-2.0. Mathlib copyright and author notices remain embedded in source contexts. See `LICENSE` and `NOTICE`.

## Release Manifest

The content-addressed public/private boundary and immutable asset references are in [`release-manifest.json`](./blob/main/release-manifest.json).