All projects
HVCap
End-to-end post-training of Qwen2.5-VL-7B: SFT cold-start followed by reinforcement learning.
- Type
- Human-centric video captioning via Factorized GRPO
- Context
- Stealth Startup
- Year
- 2025
Highlights
- Built a factorized generative ordinal reward model (Soft-Ordinal GenRM) that reuses the VLM’s LM head to grade captions on human appearance, action, background, and camera, trained on ~300k human ordinal labels.
- Developed Factorized GRPO, a Dr-GRPO-based optimizer with principled length control, running ~1.28M on-policy rollouts on an 8×H200 cluster with a decoupled reward-serving service.
- Curated HVTraining (200k examples) through a human-in-the-loop pipeline and released HVBench (3k+ gold captions); the post-trained 7B model surpasses Qwen2.5-VL-32B and InternVL3-38B and rivals Gemini-2.5 on the targeted dimensions.