README.md

---
license: mit
task_categories: [text-generation]
language: [en]
tags: [rl, confidence-estimation, calibration, math, numina]
size_categories: [1K<n<100K]
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# rl-value-confidence-train-math

**Math-domain** training set (~10K) for a confidence / correctness estimator, sampled from the
**numina** portion of [PRIME-RL/Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data).
Pairs with [`YangyiYY/rl-value-eval-math`](https://huggingface.co/datasets/YangyiYY/rl-value-eval-math)
(eval) and a 50K RL-training split from the same pool (disjoint).

Sampled **~uniformly across the 6 numina sub-sources** (cn_k12, synthetic_math, olympiads,
synthetic_amc, aops_forum, amc_aime), **deduplicated by problem text**, and **disjoint** from the
RL-train and eval splits (0 problem-text overlap).

## Schema (our recipe)
`data_source="math_box"` · `prompt=[{system: "Please reason step by step, and put your final answer
within \boxed{}."}, {user: <problem>}]` · `ability="math"` · `reward_model.ground_truth` (numina
answer) · `sub_source` (numina origin: cn_k12/synthetic_math/olympiads/synthetic_amc/aops_forum/amc_aime)
· `extra_info` (id, sub_source, split).