Eval code: https://github.com/Liquid4All/ifstruct IFStruct is a benchmark for structured-output compliance: can a model produce valid JSON/YAML that follows a requested schema, when the requirements are phrased the many different ways real users phrase them? It is scored without constrained decoding, and only the structure is judged — not content quality, extraction accuracy, or reasoning — so the signal is isolated to output validity and schema following. Each… See the full description on the dataset page: https://huggingface.co/datasets/kelexine/ifstruct-v1.0.