osim-8b-mid

OSim-8B-Mid

OSim-8B-Mid is the midtrained checkpoint of OSim (OdysSim), a foundation model for human behavior simulation — trained to imitate the human / user side of interactions rather than to behave as a helpful assistant.

It is Qwen/Qwen3-8B-Base midtrained on the OdysSim corpus: 62 public behavioral datasets unified into a common conversational format (~21.4M interactions, ~10B tokens), organized along the five Soul capability axes — CONV (discourse/interaction), SS (social skills), COG (cognitive / mental-state reasoning), ROLE (persona, roleplay, pedagogy), and EVAL (judgment / preference). Midtraining shifts the base model's prior toward the human-side distribution, avoiding the verbose, agreeable "assistant register" induced by helpfulness post-training.

This is the mid stage only (no task-specific RL or expert distillation). For the post-trained instruct model, see sunweiwei/OSim-Inst-8B.

Intended use

Simulating the human/user side of conversations — user simulation for agent evaluation, social simulation, and persona / role-play. The model is conditioned on a social-context system prompt describing who is speaking (role, goal, background, conversational style); given the other party's turns it generates the next human turn.

Results (from the paper)

On the held-out OdysSim evaluation, OSim-8B-Mid attains PPL 4.95 / BLEU 26.72 — the lowest perplexity among 8B behavioral-simulation baselines (UserLM-8B 8.38, CoSER-8B 8.77, Llama-3.1-8B 10.04).

Training

• Base: Qwen3-8B-Base
• Stage: midtraining (SFT on the OdysSim corpus, ~10B tokens / 4,500 steps)
• Optimizer: AdamW, peak LR 1e-5, 16K-input / 8K-response context, global batch 1,024, bf16, FSDP-2.

Citation

If you use this model, please cite the OdysSim paper (Building Foundation Models for Human Behavior Simulation). Code: https://github.com/sunnweiwei/OdysSim