Spaces:

---

TIGER-Lab

/

MMEB-Leaderboard

like 93

Running

Here, we present the results of the GVE-7B.

• Model: GVE-7B
• Paper: Towards Universal Video Retrieval: Generalizing Video Embedding via Synthesized Multimodal Pyramid Curriculum
• Evaluation: mostly from Qwen3-VL-Embedding

NOTE that:

• All Public Data: GVE-7B has been trained on 13M publicly available retrieval data (including 1.55M synthesized data based on public videos), providing detailed reproduction configurations
• Fully Zero-shot: No in-domain data for MMEB-V2-Video datasets is included in the training stages of the GVE series
• Retrieval Data Only: We do not utilize any video QA, classification, or grounding data
• Test-time Scaling: We report the best performance among the tested results from different test-time configurations of context length