README.md

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

# rl-value-confidence-train-code

**Code-domain** training set (~4.9K) for a confidence / correctness estimator, sampled from the raw
competitive-programming datasets **Codeforces (MatrixStudio)**, **BAAI/TACO**, and
**deepmind/code_contests**. Pairs with
[`YangyiYY/rl-value-eval-code`](https://huggingface.co/datasets/YangyiYY/rl-value-eval-code) (eval)
and a ~25K RL-training split from the same pool (disjoint).

Built by taking ALL distinct usable problems, **deduped by problem text GLOBALLY** (these datasets
heavily re-share Codeforces problems; small-source-first priority cf→cc→taco), split per source
50:10:5 into train/conf/eval (**0 problem-text overlap** across splits).

## Grader / ground_truth
`data_source` is `codeforces` / `taco` / `codecontests` — each routes to the shared reference
grader (`reference/`, apps_execution) both for the RL reward and these labels, so they are IDENTICAL.
`reward_model.ground_truth` = `input_output` JSON with **the full hidden test suite, capped at 50
test cases/problem** (problems whose gt still exceeds 1MB — a single giant test — are dropped, as
they can't be graded within the 60s/problem cap).

## Schema
`data_source` · `prompt=[{system}, {user: problem}]` · `ability="code"` · `reward_model.ground_truth`
· `sub_source` (codeforces/codecontests → rating band; taco → difficulty) · `extra_info` (id, rating, sub_source).