M2KR-WWW2025-Challenge Dataset

A multimodal retrieval dataset for image-to-document and image+text-to-document matching, used as Task 2 of Track 1 (Multimodal Document Retrieval Challenge) at the EReL@MIR Workshop, The Web Conference (WWW) 2025.

Ground truth released. Unlike the original held-out challenge release, this dataset includes the evaluation ground truth in the challenge_answers config (see below). This is a full mirror of Jingbiao/M2KR-Challenge, which remains available unchanged for backward compatibility.

Dataset Overview

• challenge_data: query data with images and optional text questions (6,415 samples; ground-truth fields blank, as in the original challenge release)
• challenge_passage: document collection with textual passages and web screenshot paths (47,318 passages)
• challenge_answers: the evaluation ground truth — positive passage IDs per query (6,415 queries)

Dataset Structure

challenge_data (6,415 rows)

Columns:

• img_path: Image filename (string)
• instruction: Task instruction for description generation
• question: Optional text query (~47% populated; if present the task is image+text-to-document retrieval, else image-to-document)
• question_id: Unique identifier (string)
• pos_item_ids: Ground-truth passage IDs — blank in this config (see challenge_answers)
• pos_item_contents: Ground-truth passage contents — blank in this config

Task types:

1. Image-to-Document Retrieval: when question is empty (image query)
2. Multimodal Retrieval: when question contains text (image + text query)

challenge_passage (47,318 rows)

Columns:

• passage_id: Unique passage identifier (string), e.g. A27558, Q5247669
• passage_content: Textual description (image description + structured entity details: birth/death dates, occupations, locations, etc.)
• page_screenshot: Associated web-screenshot filename (string)

For the retrieval task, retrieve the corresponding passage from the 47K-passage pool for each sample in challenge_data.

challenge_answers (6,415 rows) — ground truth

Columns:

• question_id: matches challenge_data.question_id (string)
• pos_item_ids: list of ground-truth passage_ids for that query (matches challenge_passage.passage_id)

from datasets import load_dataset
queries = load_dataset("Jingbiao/M2KR-WWW2025-Challenge", "challenge_data")["train"]
answers = load_dataset("Jingbiao/M2KR-WWW2025-Challenge", "challenge_answers")["train"]
# join on question_id to attach the ground truth
gt = {a["question_id"]: a["pos_item_ids"] for a in answers}

Evaluation metric. Mean of Recall@1, Recall@3, Recall@5. Recall is computed against the full positive set, so a query with multiple positives requires all of them in the top-k to score 1.0. Most queries have a single positive; 601 have more than one. The PreFLMR ViT-G baseline scores R@1/3/5 = 19.5 / 34.0 / 41.2 (mean 31.6).

Note on passage IDs. The released pos_item_ids use the canonical challenge_passage.passage_id form (e.g. A27558, Q5247669) and join directly to the corpus. (The internal Kaggle scoring used an A-stripped form for numeric IDs; that quirk is normalised away here.)

Images

To avoid duplicating ~121 GB of data, the image archives are not re-hosted here; they live in the original Jingbiao/M2KR-Challenge repo and are referenced by filename from challenge_passage.page_screenshot and challenge_data.img_path:

• Web screenshots (passage images), in Web_Image.zip.001, .002, .003
• Query images, in query_images.zip

from huggingface_hub import hf_hub_download
# fetch the image archives from the original data repo
for part in ["Web_Image.zip.001", "Web_Image.zip.002", "Web_Image.zip.003", "query_images.zip"]:
    hf_hub_download("Jingbiao/M2KR-Challenge", part, repo_type="dataset", local_dir="images/")

References

• Paper: https://arxiv.org/abs/2402.08327
• Project Page: https://preflmr.github.io/
• FLMR implementation: https://github.com/LinWeizheDragon/FLMR
• Original challenge release: https://huggingface.co/datasets/Jingbiao/M2KR-Challenge

Citation

If this dataset helped your research, please cite PreFLMR:

@inproceedings{lin-etal-2024-preflmr,
    title = "{P}re{FLMR}: Scaling Up Fine-Grained Late-Interaction Multi-modal Retrievers",
    author = "Lin, Weizhe and Mei, Jingbiao and Chen, Jinghong and Byrne, Bill",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.289",
    pages = "5294--5316",
}