CoolFace
Datasetpublic

CL-From-Nothing/RLVE-Qwen3-4B-Thinking-2507-Pass8-Rollouts

RLVE teacher rollouts — Qwen3-4B-Thinking-2507 (pass@8) Teacher rollouts for on-policy distillation on the RLVE environment suite. Teacher / sampler: Qwen3-4B-Thinking-2507 Source prompts: RLVE train split — 9000 questions across 18 environments (counting / combinatorics / optimization tasks) Sampling: 8 samples/question (pass@8) = 72000 records, temperature 1.0 (sample.sh default 0.7 -> here T per run), max 16384 new tokens Rewards: recomputed offline with the RLVE-Eval Gym… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Qwen3-4B-Thinking-2507-Pass8-Rollouts.

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes22downloads
Dataset Card

RLVE teacher rollouts — Qwen3-4B-Thinking-2507 (pass@8)

Teacher rollouts for on-policy distillation on the RLVE environment suite.

  • —Teacher / sampler: Qwen3-4B-Thinking-2507
  • —Source prompts: RLVE train split — 9000 questions across 18 environments (counting / combinatorics / optimization tasks)
  • —Sampling: 8 samples/question (pass@8) = 72000 records, temperature 1.0 (sample.sh default 0.7 -> here T per run), max 16384 new tokens
  • —Rewards: recomputed offline with the RLVE-Eval Gym verifier (reward in [-1, 1]: +1 correct, -0.5 wrong answer, -1 wrong/missing format). Teacher accuracy (reward>0): 23153 / 72000 = 32.2%.

Record schema (one JSON object per line)

fieldmeaning
indexquestion index (0-8999)
sample_idsample id within the question (0-7)
data_source"rlve"
promptthe problem statement
responsethe teacher's generated solution (full chain-of-thought)
rewardsGym-verifier reward in [-1, 1] (re-scored)
metadataJSON string: environment name + config (for the verifier)
answerground-truth answer