Training/eval data for the Diagnosis-Advantage RLVR recipe (verl GRPO): given a
correct-but-slow Python solution to a competitive-programming problem, the model
must produce a faster, still-correct optimization. Each problem appears as two
prompt variants sharing one uid (route-B paired rollouts):
variant A — diagnose then optimize (emit a
… block, then optimized code)
variant B — optimize directly (no diagnosis) — the… See the full description on the dataset page:
https://huggingface.co/datasets/DONG19/diagnosis-advantage-grpo.
