MMEB-v2 (VLM2Vec eval bundle)

Portable Hugging Face bundle of MMEB-V2 evaluation queries with embedded query images, ground-truth metadata, and retrieval predictions from:

• Eval run: eval_Qwen2vl2B_merged_step5k_benchmark
• Local eval data root: vlm2vec_eval under VLM2Vec (/public/hk-research/users/songliyue/code/VLM2Vec/data/vlm2vec_eval)

Structure

• 24 configs (one per benchmark task): image/*, video/*, visdoc/*
• Each row = one query with query_text, query_images, candidates, labels, and model_prediction ranking

Load

from datasets import load_dataset
dd = load_dataset("Fox4Song1/MMEB-v2")
ds = dd["image__A-OKVQA"]  # example config name

Config names use {modality}__{task} (double underscore).