Judge inference outputs (selection decisions over paired candidate code) on BigCodeBench,
produced with forcing mode for SFT warmup. Migrated 2026-04-24 from 2 separate repos
(t2ance/selection_bcb_sft_warmup_forcing_30b_{train,test}) into a unified config+split structure.
Config = judge/candidate-source model; split = train/test.
Predecessor repos were deleted on 2026-04-25 after byte-row-level equivalence was verified (row count + schema + per-row SHA256… See the full description on the dataset page:
https://huggingface.co/datasets/t2ance/selection-bcb-sft-warmup.
