ThakiCloud/korean-contract-derivation-bench
Preview • Updated • 46
Benchmarks and reproduction material for re-measuring the numbers on the model cards. Kept separate from the release collections.
Note Korean contract RAG — hallucination and instruction-following evaluation
Note Requantization rebuild noise floor — read this before attributing any difference to a recipe. Ships with a 6-page paper (paper/, CC BY 4.0)
Note Korean dialect benchmark, 1,000 items across 5 regions — no AI Hub derivatives are distributed (item ids hashed, marker lexicon regenerated)
Note ChainCheck — separates chain sensitivity from edit sensitivity in evidence-sufficiency scorers