dna_hallucination

Evaluation set for hallucination resistance: does the model abstain or correct the question instead of fabricating?

Composition

100 Korean prompts with reference answers, 20 in each of five categories (category):

category	what it tests
false_premise	the question presupposes something false; the model should correct it
true_premise	a matched control with a true premise; the model should answer
nonexistent	the subject does not exist; the model should say so
unknowable	no reliable answer exists; the model should abstain
grounding	the answer must stay within the information given

pair_id links false/true-premise pairs; difficulty and mode are present where defined.

Scoring

Each response is judged against target. Report the overall score and the per-category breakdown.

Part of the DNA evaluation suite (dnotitia/dna_*). Responses are judged by an LLM against a fixed rubric; the reported score is the 15% trimmed mean over 15 runs, together with the pass rate. Evaluation code: dnotitia/dna-eval.