AgentPlays
Microsoft's UserRepGen-4B turns a chat log into a user profile for testing assistants
UserRepGen-4B, an MIT-licensed fine-tune of Qwen3.5-4B, reads a multi-turn user-assistant conversation and writes a JSON profile with a task description and a behavior description (communication style, how constraints come out, how feedback is given). The profile is meant to drive a separate LLM-based user simulator for testing assistants; the model doesn't write user turns itself. English only. The model card flags a privacy risk: in Microsoft's tests the behavior descriptions helped match conversations to the same user, so they are not anonymized data. It says to use only conversations you're authorized to process and to remove identifying details first.
How you’d use it: Turn a real support chat into a simulated user to regression-test your assistant.