seokwon99/Multi3IR

Multi³IR-Bench

Multi-perspective, multi-domain, multi-modal information retrieval benchmark (paper, code, project page). 104.9K Stack Exchange questions (72K train / 32K validation / 1.0K test) with 521.7K perspective descriptions (4.98 per query) and judged gold documents per perspective, over a 1.01M-document corpus of C4 web text and web images. This release ships the train split (train + validation questions together) and the 1.0K test split.

config	split	rows	columns
queries	train / eval	103,877 / 1,039	uid, query, perspectives
gold	train / eval	632,558 / 42,389	uid, perspective_idx, summary, doc_id, modality, url, domain (train), rank + text (eval; text inline for text docs)
corpus	text / image	457,168 / 346,829	doc_id, url — the benchmark corpus (every document that is gold for some question)

Documents are referenced by id and source url, not redistributed: C4 passages (c4-<file>-<line>) can be recovered from allenai/c4, images from their url. In the code repository, python -m multi3ir.data.hub pull restores the benchmark layout and python -m multi3ir.data.build_corpus materialises the corpus.

from datasets import load_dataset
gold = load_dataset("seokwon99/Multi3IR", "gold", split="eval")

Sources: Stack Exchange (CC BY-SA 4.0), C4 (ODC-BY), images © their owners (research use). Perspectives were written by GPT-5-mini and filtered for faithfulness (Bespoke-MiniCheck-7B) and uniqueness (all-MiniLM-L6-v2, cosine > 0.8); document support was judged by Qwen3-VL-30B-A3B-Instruct (agreement 0.90 with human labels on fully).

@article{song2026multi3ir,
  title={MULTI3IR: A Benchmark for Multi-perspective Multi-domain Multi-modal Information Retrieval},
  author={Song, Seokwon and Kim, Sohyeon and Kim, Gunhee},
  journal={arXiv preprint arXiv:2608.30949},
  year={2026}
}