Model reports

NAFBench

原始来源英文简介

Solver-certified normal logic programs verbalized into natural language, each labelled under four negation semantics (SLDNF, well-founded, and credulous and skeptical stable-model reasoning) that can disagree on the same instance. A score without the semantics named is not comparable. Models are also sensitive to rule order. The leaderboard's hidden test split (Hugging Face qbao775/naf-bench, as described by the repository README at revision 83ede91) has 495 rows drawn from only 77 programs, so the effective sample is the program, not the row. The arXiv v1 appendix describes an earlier 385-prompt evaluation set over 77 programs, so a paper number and a leaderboard number are not on the same split.