Spaces:

---

TIGER-Lab

/

MMEB-Leaderboard

like 104

Running

Description

This PR adds the evaluation scores for Qwen3-VL-embedding-Turbo1.9B_v1, a lightweight multimodal embedding model developed at Urock.

This model is derived from Qwen3-VL-Embedding and optimized through structural pruning and layered distillation.

Key Highlights:

• Model Size: 1.9B (10% reduction compared to the base model)
• Performance: High efficiency with ~10% faster inference latency while maintaining competitive performance on MMEB tasks, especially in VisDoc (Visual Document Understanding).
• Language Support: English (en) and Korean (ko).

Metadata:

• Model Name: Qwen3-VL-embedding-Turbo1.9B_v1
• Model URL: https://huggingface.co/zxcv1245/Qwen3-VL-embedding-Turbo1.9B_v1
• Model Size: 1.9 (Billion parameters)
• Data Source: Self-Reported (evaluated using the official MMEB pipeline)

I have uploaded the generated JSON file to the scores/ directory as per the instructions.

Please review this submission and let me know if any further information is required.