I find smaller models get confused with braces for JSONish output. Has anyone looked if yaml output format produces more reliable output?

Interesting, which models are you usinhrg?

Happy Thanksgiving! Minimax M1 and M2, Kimi-K2 (thinking), and Cohere models were the most inconsistent. Other models struggled occasionally as well, but these were noticeably less stable—especially when responding in non-English languages.