Per-question teacher/student reward-delta annotations for verifier-free self-self-distillation,
computed on the sky_work_math subset of
PrimeIntellect/SYNTHETIC-2-RL with
Qwen/Qwen3-4B.
For each problem we draw k=8 rollouts in thinking-on (teacher) and thinking-off (student) modes at
identical sampling (temperature 0.7 / top_p 0.8), grade each against the ground truth, and record the
per-mode expected reward and their difference (delta = R_teacher -… See the full description on the dataset page:
https://huggingface.co/datasets/rdavion/self-self-distillation.
