lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-budget8192

Matched no-conftest RLVR study 20260909-budget8192 Retired before study training. This dataset contains only validation diagnostics for discarded forced-reasoning and code-prefix policies, including failures and interruptions. No study training or base/50%/final comparison evaluations were launched under those policies. They are excluded from the active completion-reward study. All available diagnostic records are preserved losslessly below. Lossless research records, grouped by… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-budget8192.