hanjianghu/VLESA-Qwen3-VL-2B-Safety-Filter

VLESA Safety Filter — Qwen3-VL-2B (GRPO fine-tuned)

This is the VLESA safety filter checkpoint from the paper:

VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring
Hanjiang Hu, Yiyuan Pan, Jiaxing Li, Xusheng Luo, Alexander Robey, Na Li, Yebin Wang, Changliu Liu
arXiv:2606.03954 | GitHub

Model Description

This model is a Qwen3-VL-2B checkpoint fine-tuned via Group Relative Policy Optimization (GRPO) on the EgoSafety dataset. It serves as the safety filter component of the VLESA pipeline: given egocentric video frames (pre-action, point-of-no-return, and post-action), it predicts whether a human action is safe or unsafe and provides a natural-language rationale.

Training Details

• Base model: Qwen3-VL-2B-Instruct
• Training method: GRPO (Group Relative Policy Optimization) via EasyR1
• Training data: EgoSafety — 9,596 train / 1,202 validation / 1,183 test samples derived from Ego4D
• Checkpoint: global_step_148

Usage

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "hanjianghu/VLESA-Qwen3-VL-2B-Safety-Filter",
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("hanjianghu/VLESA-Qwen3-VL-2B-Safety-Filter")

For full pipeline usage with the VLESA agent, see the GitHub repository.

Citation

@article{hu2026vlesa,
  title={VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring},
  author={Hu, Hanjiang and Pan, Yiyuan and Li, Jiaxing and Luo, Xusheng and Robey, Alexander and Li, Na and Wang, Yebin and Liu, Changliu},
  journal={arXiv preprint arXiv:2606.03954},
  year={2026}
}