File size: 1,747 Bytes
db11f26 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | ---
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
base_model: Qwen/Qwen3.5-9B
datasets:
- OS-Copilot/OS-Shepherd-100K
language:
- en
tags:
- computer-use
- reward-model
- trajectory-evaluation
- multimodal
---
# OS-Shepherd-9B
OS-Shepherd-9B is an open multimodal reward model for judging computer-use agent trajectories. Given a task instruction, screenshots, and the agent's reasoning and actions, it determines whether the task was completed and returns a reasoned `SUCCESS` or `FAIL` verdict.
It is fine-tuned from [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) on [OS-Shepherd-100K](https://huggingface.co/datasets/OS-Copilot/OS-Shepherd-100K) using SFT followed by GRPO, with the RL stage focused on reducing false-success judgments.
## Results
| Benchmark | Accuracy | Fail recall |
|---|---:|---:|
| OSReward | 86.1 | 86.0 |
| OSReward-Hard | 60.2 | 57.6 |
Results use the fixed judging protocol described in the [OSReward paper](https://arxiv.org/abs/2607.28609).
## Usage
Use the canonical prompt and trajectory format from the [OSReward repository](https://github.com/OS-Copilot/OSReward). A recent Transformers, vLLM, or SGLang version with Qwen3.5 multimodal support is required.
This model is intended for trajectory evaluation, data filtering, and reward-model research. It is not a computer-control policy and may still miss fine-grained visual failures, especially on hard cases.
## License
Apache License 2.0. See `LICENSE`.
## Citation
```bibtex
@article{sun2026osreward,
title={OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models},
author={Sun, Qiushi and others},
journal={arXiv preprint arXiv:2607.28609},
year={2026}
}
```
|