Wouldn’t it be better to use a grammar in the token sampler? Tuning is fine, but doesn’t guarantee a syntactical correct structured output. But if the sampler is grammar aware it could.