Datasets:

YangyiYY
/

rl-value-confidence-train

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

rl-value-confidence-train

A held-out training set for a confidence / correctness estimator (~5.7k prompts), paired with YangyiYY/rl-value-eval (evaluation) and the nct-ppo actor/critic models. Each row is a (prompt, ground_truth) in the exact verl RL schema, so the same rule-based verifiers score it (math: \boxed{} grading; code: stdin/stdout or function test cases).

Built to be disjoint from both the RL training data and the calibration eval sets (enforced by md5 of the normalized problem text — 0 overlap), at matched difficulty where the pools allow.

Composition (5728 rows)

data_source	n	source	difficulty
`math_box` (numina)	2000	AI-MO/NuminaMath-CoT `train`, boxed answers	same distribution as RL training (sub-datasets: cn_k12, synthetic_math, olympiads, orca_math, …)
`taco`	2998	BAAI/TACO `train`	sampled to match the RL-training taco difficulty mix (EASY/MEDIUM/HARD/…)
`codecontests`	655	deepmind/code_contests (train+valid+test)	cf_rating 1201–1400
`codeforces`	75	MatrixStudio/Codeforces-Python-Submissions	rating 1201–1400

Note on code difficulty: the RL training + eval sets consumed all Codeforces/CodeContests problems at rating ≤1200 (the training band). This held-out set therefore takes the closest available band (1201–1400) for code — one notch above training — since ≤1200 has zero disjoint problems left. numina and taco are at matched training difficulty.

Schema

data_source · prompt (list of {role, content} chat messages) · ability (math/code) · reward_model.ground_truth (rule verifier target) · extra_info (id, rating, difficulty, sub_source, split="confidence_train").

Intended use

Train a correctness/confidence estimator (e.g. a value head or probe) on rollouts from the frozen nct-ppo actor, then evaluate calibration (ECE/AUROC) on the disjoint rl-value-eval set.