Публикация

Cheering myself up from [REDACTED] by watching models debug other models: > Llama-Guard-3-1B is technically working and producing the right output format, but it's giving false positives — it flagged "How do I bake bread?" as unsafe,