Comment by anthonyiscoding

That explains why gpt-oss wasn't working anywhere near as well for me as other similarly and smaller sized models. gemma3 27b, 12b, and phi4 (14b?) all significantly outperformed it when transforming unstructured data to structured data.