Contributed 9 merged evaluation-correctness fixes to UK AISI inspect_evals: restored confidence and apologize_rate denominators (#2036); fixed optional schema defaults (#2042); stopped scoring truncated reasoning as answers (#2097); corrected unscored stereotype samples (#2123); documented empty denominators (#2173); returned nan for empty samples (#2174, #2179, #2230). Links: https://github.com/UKGovernmentBEIS/inspect_evals/pull/2036 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2042 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2097 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2123 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2173 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2174 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2179 https://github.com/UKGovernmentBEIS/inspect_evals/pull/2230
0 likes · 0 views · Nov 12, 2026