Search-Gen-V-4b

Search-Gen-V is not only capable of efficiently determining whether text segments meet various evaluation criteria, but it can also output its reasoning process, thereby enhancing the transparency and interpretability of the evaluation results. At the same time, it supports batch verification across multiple evaluation criteria and significantly improves reasoning efficiency and resource utilization by constraining the generated tokens, maintaining high accuracy while reducing computational costs.

---

🧠 Model Details

• Base model: [Qwen3-4B]
• Fine-tuning method: [SFT+DAPO]
• Training framework: [VeRL]

---

🚀 Usage Example

MODEL_PATH=$1
python3 -m sglang.launch_server \
    --model-path ${MODEL_PATH} \
    --tp-size 8 \
    --max-running-requests 160 \
    --cuda-graph-max-bs 160 \
    --mem-fraction-static 0.8 \
    --reasoning-parser qwen3 \
    --tool-call-parser llama3 \
    --trust-remote-code \
    --host 0.0.0.0 \
    --port 8000

---

Citation

@article{ma2025searchgenv,
  title={AN EFFICIENT RUBRIC-BASED GENERATIVE VERIFIER FOR SEARCH-AUGMENTED LLMS},
  author={Ma, Linyue and Xu, Yilong and Long, Xiang and Zheng, Zhi},
  journal={arXiv preprint arXiv:2510.14660},
  year={2025},
  url={https://arxiv.org/abs/2510.14660}
}