---
license: mit
task_categories: [text-generation]
language: [en]
tags: [rl, confidence-estimation, calibration, math, numina, aime]
size_categories: [1K<n<10K]
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# rl-value-eval-math

**Math-domain** evaluation set (~5.1K) for confidence / correctness estimation. Pairs with
[`YangyiYY/rl-value-confidence-train-math`](https://huggingface.co/datasets/YangyiYY/rl-value-confidence-train-math)
(confidence-train) and a 50K RL-training split (disjoint).

Bulk of the set is sampled ~uniformly across the 6 **numina** sub-sources (from
[PRIME-RL/Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data)),
deduplicated by problem text and disjoint from the RL-train / confidence-train splits. It also
includes two held-out competition sets:
- **AIME 2024** (30 problems, [math-ai/aime24](https://huggingface.co/datasets/math-ai/aime24))
- **AIME 2025** (30 problems, [math-ai/aime25](https://huggingface.co/datasets/math-ai/aime25))

## Composition (`sub_source`, 5058 rows)

| sub_source | n | source |
|---|---|---|
| synthetic_math / cn_k12 / olympiads / synthetic_amc | 965 each | numina |
| aops_forum | 949 | numina |
| amc_aime | 189 | numina |
| **aime24** | 30 | math-ai/aime24 |
| **aime25** | 30 | math-ai/aime25 |

## Schema (our recipe)
`data_source="math_box"` for all rows (routes to the shared math grader) · `prompt=[{system: "Please
reason step by step, and put your final answer within \boxed{}."}, {user: <problem>}]` ·
`ability="math"` · `reward_model.ground_truth` (numina answer / AIME integer) · `sub_source`
(numina origin or `aime24`/`aime25`) · `extra_info` (id, sub_source, split).