SCOPE-SFT-9B

SCOPE-SFT-9B

SCOPE-SFT-9B is a computer-use agent (CUA) trained under the SCOPE (Safety and Capability Optimization for Policy Execution) framework to balance task-execution capability with safety-aware decision-making. The model is initialized from Qwen3.5-9B and jointly fine-tuned on capability demonstrations, safe-continuation trajectories, and explicit-refusal trajectories.

In our evaluation, SCOPE-SFT-9B achieves a 49.72% task success rate on OSWorld and a 66.30% attack-avoidance rate on OS-BLIND, corresponding to a capability-safety harmonic mean of 56.83%.

SCOPE-SFT-9B is trained on SATraj-OS using joint supervised fine-tuning. Capability demonstrations teach the model to complete benign desktop tasks, while safe-continuation and explicit-refusal trajectories teach it to respond appropriately when a request or execution environment presents a safety risk. SCOPE-SFT-9B also serves as the initialization checkpoint for SCOPE-RL-9B.

Links

• Paper: Beyond Task Completion: Training Capable and Safe Computer-Use Agents
• Training dataset: SATraj-OS
• Data and safety framework: Safactory
• Model collection: SCOPE

Quick Start

Install vLLM:

pip install -U vllm

Launch an OpenAI-compatible inference server:

vllm serve k4ng/SCOPE-SFT-9B \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --data-parallel-size 2 \
    --trust-remote-code \
    --served-model-name scope-sft

Results

All values are percentages, and ↑ indicates that higher is better. (H) is the harmonic mean of the OSWorld task success rate and the OS-BLIND attack-avoidance rate.

License

This model is subject to the license terms of its base model, Qwen3.5-9B. Licensing information for the code and training data is available in the corresponding repositories.

Citation

If you use SCOPE-RL, SCOPE-SFT, SATraj-OS, or SCOPE-Gen, please cite:

@misc{kang2026scope,
  title  = {Beyond Task Completion: Training Capable and Safe Computer-Use Agents},
  author = {Zeyu Kang and Zhenyun Yin and Yang Zhang and Shan He and Shanzhe Lei and Yanjiu Zhong and Xinquan Chen and Xuhong Wang},
  year   = {2026}
}