36 lines
2.2 KiB
Markdown
36 lines
2.2 KiB
Markdown
|
---
|
|||
|
|
license: mit
|
|||
|
|
base_model:
|
|||
|
|
- Qwen/Qwen3-8B
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
tags:
|
|||
|
|
- human-simulation
|
|||
|
|
- user-simulation
|
|||
|
|
- behavioral-modeling
|
|||
|
|
- social-simulation
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# OSim-8B (post-trained, text)
|
|||
|
|
|
|||
|
|
**OSim-8B** is the post-trained (text) 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 the text counterpart of the midtrained [`cmu-lti/osim-8b-mid`](https://huggingface.co/cmu-lti/osim-8b-mid): [`Qwen/Qwen3-8B`](https://huggingface.co/Qwen/Qwen3-8B) midtrained on the OdysSim corpus (62 behavioral datasets along the five **Soul** axes — CONV/SS/COG/ROLE/EVAL) and then post-trained (task-specific RL + expert consolidation).
|
|||
|
|
|
|||
|
|
*(Mirror of `sunweiwei/OSim-8B`, the text post model. For the VL variant see `sunweiwei/OSim-Inst-8B`.)*
|
|||
|
|
|
|||
|
|
## Intended use
|
|||
|
|
|
|||
|
|
Simulating the human/user side of conversations — user simulation for agent evaluation, social simulation, persona / role-play. Conditioned on a "social-context" system prompt (who is speaking: role, goal, background, style); given the other party's turns it generates the next human turn.
|
|||
|
|
|
|||
|
|
## Results
|
|||
|
|
|
|||
|
|
Evaluated **out-of-distribution** as the user simulator in the τ-USI agentic benchmark (τ-bench airline+retail, 165 tasks, fixed GPT-5.2 agent), OSim-8B reaches **USI 75.6** — the **best behavioral / specialized user simulator**, surpassing same-size general instruct models and every prior specialized simulator (CoSER-8B 67.2, UserLM-8B 62.0). It is distinctively human-like in **reactivity** (Sørensen–Dice D4 ≈ 93, matching the human inter-annotator level) and in **outcome calibration** (best ECE among compared models), with essentially none of the long-horizon agentic failure modes (timeouts/perseveration) seen in non-behavioral baselines.
|
|||
|
|
|
|||
|
|
## Training
|
|||
|
|
|
|||
|
|
- **Base:** Qwen3-8B
|
|||
|
|
- **Stages:** midtraining on the OdysSim corpus → task-specific reinforcement learning + expert consolidation.
|
|||
|
|
|
|||
|
|
## Citation
|
|||
|
|
|
|||
|
|
If you use this model, please cite the OdysSim paper (*Building Foundation Models for Human Behavior Simulation*). Code: https://github.com/sunnweiwei/OdysSim
|