Per-prefix aggregate statistics for run
fineproofs_all_qwen35_9b_direct2phase_m32_20260730.
Each row represents one collected response prefix and summarizes exactly 32 fresh
Qwen3.5-9B continuations graded by GPT-OSS-120B against the problem rubric. Every
continuation and endpoint reward is reconstructed exactly as clamped
points / max_points; the rounded journal grade is used only for a consistency
audit with absolute tolerance… See the full description on the dataset page:
https://huggingface.co/datasets/asingh15/fineproofs-branch-rollout-distribution.
