Browse skills

IFEval

v1writing.structured

Google's Instruction-Following Eval — 25 verifiable-instruction types scored by rule validators (no LLM judge). 100-task subsample of the public split, held constant to enable version-over-version comparison.

Scored by ifeval over 100 tasks. Methodology: see BENCHMARKS.md in the source repo.

RankSkillVersionAccuracyCost / successMedian latencyRobustness
category median
No completed runs yet — skills in this category are still being scored.