MMhops-KB

MMhops-KB contains the minimal text and image retrieval databases used by MMhops-R1. The JSONL row ID is the corresponding row ID in its FAISS index.

Files

MMhops-KB/
├── text/
│   ├── e5_Flat.index
│   └── text_corpus.jsonl
└── image/
    ├── CLIP_Flat.index
    └── image_corpus.jsonl

• text/text_corpus.jsonl contains 904,790 chunked Wikipedia passages as {"id": "0", "contents": "title\npassage"}.
• image/image_corpus.jsonl contains 96,110 indexed entities as {"id": 0, "title": "...", "entity_id": "Q...", "image_path": "..."}.

Retrieval configuration

The text index uses intfloat/e5-base-v2, 768-dimensional mean-pooled embeddings, and FAISS Flat inner-product search. The image index uses OpenAI CLIP ViT-L/14@336px, 768 dimensions, and Flat inner-product search. Each indexed image vector combines normalized image and title-text features with weights 0.5 and 0.5.

Model weights are external dependencies and are not duplicated here.

Upstream images

Image bytes are not redistributed. image_path points into wikipedia_images_full/ in the gated OVEN snapshot ychenNLP/oven/all_wikipedia_images.tar. After accepting the upstream access conditions, download it with:

hf download ychenNLP/oven all_wikipedia_images.tar \
  --repo-type dataset --local-dir .
tar -xzf all_wikipedia_images.tar

The archive is gzip-compressed despite its .tar suffix. Raw images are not needed to serve the prebuilt image index; they are only needed to inspect or rebuild the database.

Sources and licenses

The passage corpus is derived from English Wikipedia and remains subject to Wikipedia's applicable reuse terms. The image snapshot remains governed by its upstream OVEN/Wikimedia terms. The MMhops-R1 software license does not replace those upstream content licenses.

Citation

@misc{zhang2025mmhopsr1,
  title={MMhops-R1: Multimodal Multi-hop Reasoning},
  author={Tao Zhang and Ziqi Zhang and Zongyang Ma and Yuxin Chen and Bing Li and Chunfeng Yuan and Guangting Wang and Fengyun Rao and Ying Shan and Weiming Hu},
  year={2025},
  eprint={2512.13573},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}