SEP-100 is a synthetic, theory-neutral matched-pair benchmark for single-constraint policy transitions. The frozen benchmark contains 50 pairs / 100 items. This repository releases the labeled public half: 25 pairs / 50 items. Sealed prompts and labels are not included.
data/test.jsonl: labeled public split.
scripts/score_sep100.py: strict typed scorer.
scripts/validate_release.py: release validator.
tests/gold_predictions.jsonl:… See the full description on the dataset page:
https://huggingface.co/datasets/John6666/sep-100-public.
