Benchmarks & research
PolicyLM-1.7B reads a message beside your moderation rules and returns a score for each category, without writing a reply.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Musubi's v1.2 helper allows sixteen custom categories. Their rules and the message share 2,048 tokens. The Apache-2.0 model handles text only, one message at a time; it cannot read an image or supply a written reason for an appeal. Musubi reports its own speed and language tests. The released model card says it has not been tested on live traffic.
For each category, describe a violation, a non-violation and any exception. The source's example treats quoting a threat to report it differently from making a threat. Try such pairs before choosing the cutoff that flags a message. Do not use this model as a child-safety enforcement tool or as the only safeguard for self-harm.