PolicyShiftGuard-3B-RP-SFT

This repository releases the Stage-1 Randomized Policy SFT (RP-SFT) checkpoint for the 3B PolicyShiftGuard model.

📚 Paper | 💻 GitHub | 🏠 Project Page

RP-SFT is the first training stage in PolicyShiftGuard. It trains a Qwen2.5-VL guardrail model to read policy bundles under randomized policy identifiers and randomized policy ordering. This checkpoint is provided for reproducibility and ablation use. The final public model after the second-stage adaptation is available at PolicyShiftGuard/PolicyShiftGuard-3B.

Intended Use

Use this checkpoint when you want to reproduce the two-stage training pipeline or compare Stage-1 RP-SFT against the final BP-Adapt model.

For standard evaluation or deployment, use the final model instead:

• PolicyShiftGuard/PolicyShiftGuard-3B

Dataset

The model is trained with PolicyShiftBench supervision:

• PolicyShiftBench/PolicyShiftBench

Notes

• This is an intermediate checkpoint, not the final model reported as the main PolicyShiftGuard model.
• This checkpoint corresponds to the randomized-policy no-think Stage-1 SFT setting.
• Training-state files such as optimizer states are intentionally not included.

Citation

If you use this repository, please cite the paper:

@article{song2026policyshiftguard,
  title   = {PolicyShiftGuard: Benchmarking and Improving Policy-Adaptive Image Guardrails},
  author  = {Song, Mingyang and Xu, Luxin and Sun, Haoyu and Pan, Minzhou and Cheng, Yu and Li, Bo},
  journal = {arXiv preprint arXiv:2607.05910},
  year    = {2026}
}