File size: 2,812 Bytes
fbd89fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-nc-sa-4.0
language:
- en
task_categories:
- question-answering
tags:
- math
- competition-math
size_categories:
- 1K<n<10K
---

# CrystalMath eval (filtered)

A filtered evaluation copy of [`ycchen/Crystal-Math-Preview`](https://huggingface.co/datasets/ycchen/Crystal-Math-Preview),
config `rewritten` (competition problems whose answers are a single integer in [0, 99999], graded by exact match).

**1,790 of the 2,020 problems are kept.** 230 are left out, each listed with its reason and evidence in
`crystalmath_exclude.csv` (`excluded.csv` has id and reason only):

| reason | problems |
|---|---|
| overlap with a training set of ours (same, reworded or number-changed problem) | 77 |
| duplicate (13 exact, the rest differ only in formatting; the smallest id is kept) | 47 |
| wrong label (checked by solving and against official answers) | 28 |
| broken problem (ill-posed, no unique answer, contradictory givens, a proof, or a rewrite that changed the problem) | 23 |
| needs a figure that is not in the text | 22 |
| problem in Chinese | 21 |
| estimation scoring, garbled text, doubtful label, or outside information needed | 12 |

Labels were checked for every problem that neither gpt-oss-120b (15 tries with Python, the dataset's own pass rates)
nor 8 samples of Qwen3.5-9B ever matched, and every problem that mentions a figure, answer choices or estimation
scoring. A few wrong labels may remain among problems that are merely hard.

## Files

- `test.parquet`: all 1,790 problems, ordered by sha256 of `"0:" + id`.
- `test_1000.parquet`, `test_500.parquet`: the first 1000 / 500 of that order (random, nested: 500 inside 1000 inside
  all, and stable if the list changes).
- `blocks/b00.parquet` ... `b14.parquet`: the same order in blocks of 125 (the last 40), for splitting eval jobs; the
  500 subset is b00-b03, the 1000 subset b00-b07.

Each row: `prompt` (system: "You are a careful mathematician. Solve the following problem. Show your work, then give
the final answer inside \boxed{}."; user: the problem), `data_source` = `crystalmath`, `reward_model.ground_truth`
(the integer answer as a string), and `extra_info` (`crystalmath_id`, `domain`, `source`, `original_answer`, and the
source dataset's gpt-oss-120b `medium_pass_rate` / `high_pass_rate`).

## License and attribution

All problems, answers and difficulty labels come from Crystal-Math-Preview (Chen et al., *CrystalMath: Curating
High-Difficulty Contest Math with Contrast-Augmented Verification*), which aggregates 12 public sources, each with
its own license; the most restrictive are CC BY-NC-SA 4.0 (MathArena) and CC BY-NC 4.0 (ODA-Math-460k). This copy is
distributed under **CC BY-NC-SA 4.0** for non-commercial research; see the source dataset's `license` column for the
license of each problem.