Benchmarks & research
PolicyLM-1.7B scores a message against your moderation categories instead of writing a reply.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Musubi's v1.2 helper accepts sixteen custom categories, with their rules and the message sharing 2,048 tokens. The Apache-2.0 model reads one text message at a time. It cannot inspect an image or write a reason for an appeal. Speed and language results come from Musubi's own tests; the released model card says it has not been tested on live traffic.
Describe what violates each rule, what does not, and where exceptions apply. The source distinguishes quoting a threat to report it from making one. Test pairs like that before choosing a flagging cutoff. Do not use the model for child-safety enforcement or as the sole safeguard for self-harm.