cmu-lti/osim-8b-mid

OSim-8B-Mid: A Foundation Model for Human Behavior Simulation

OSim-8B-Mid is an 8 billion parameter model developed by cmu-lti, serving as a foundational checkpoint for human behavior simulation. Unlike typical LLMs designed as helpful assistants, OSim-8B-Mid is specifically trained to imitate the human or user side of interactions.

Key Capabilities

• Human Behavior Simulation: Trained on the OdysSim corpus, which unifies 62 public behavioral datasets (approximately 21.4 million interactions and 10 billion tokens), the model learns to generate realistic human-like responses.
• Midtraining Approach: It is a midtrained version of Qwen3-8B-Base, meaning it has undergone supervised fine-tuning on the OdysSim corpus to shift its prior towards human-side conversational distributions, effectively avoiding the verbose and overly agreeable "assistant register."
• Contextual Generation: The model can be conditioned with a social-context system prompt (describing role, goal, background, conversational style) to generate the next human turn based on the other party's input.
• Performance: Achieves a perplexity (PPL) of 4.95 and BLEU score of 26.72 on the held-out OdysSim evaluation, outperforming other 8B behavioral-simulation baselines like UserLM-8B (PPL 8.38) and Llama-3.1-8B (PPL 10.04).

Good For

• User Simulation: Ideal for evaluating AI agents by simulating realistic user interactions.
• Social Simulation: Useful for modeling and understanding social dynamics in conversational contexts.
• Persona and Role-Play: Excellent for generating human-like personas and engaging in role-play scenarios.

This model represents the midtraining stage; for a post-trained instruct version, refer to sunweiwei/OSim-Inst-8B.