Safety

FELM

FELM: a factuality-evaluation meta-benchmark of 847 ChatGPT responses (gpt-3.5-turbo) across five domains (world knowledge, science/tech, math, reasoning, writing/recommendation), each segmented into text spans with a human factuality label (factual vs. factual error) and free-text corrections. This build ingests the publicly released single-subject matrix (subject = ChatGPT, item = response segment, response = per-segment human factuality verdict) — 4,426 labeled segments. The detector-model prediction layer (GPT-4 / ChatGPT / Vicuna under raw/CoT/link methods) is not released upstream.