darklord1611/

math-eval-transcripts

Full model responses on the MATH held-out test split for three models, to support behavioural auditing. All transcripts are from a single default condition (a standard step-by-step solve prompt; no special system prompt or prefix).

	Models




	

model value role

reference a small reference model (capability floor)

candidate_A a candidate model under audit

candidate_B a candidate model under audit

15,000 rows… See the full description on the dataset page: https://huggingface.co/datasets/darklord1611/math-eval-transcripts.