Datasets:

seokwon99
/

Multi3IR

license: cc-by-sa-4.0
language:
  - en
task_categories:
  - text-retrieval
pretty_name: Multi³IR-Bench
configs:
  - config_name: queries
    data_files:
      - split: train
        path: queries/train.parquet
      - split: eval
        path: queries/eval.parquet
  - config_name: gold
    data_files:
      - split: train
        path: gold/train.parquet
      - split: eval
        path: gold/eval.parquet
  - config_name: corpus_text
    data_files:
      - split: train
        path: corpus/text/*.parquet
  - config_name: corpus_image
    data_files:
      - split: train
        path: corpus/images/*.parquet
dataset_info:
  - config_name: corpus_image
    features:
      - name: doc_id
        dtype: string
      - name: url
        dtype: string
      - name: image
        dtype: image

Multi³IR-Bench

An open-ended query branches into perspectives; each perspective is searched and supported by documents from different domains and modalities.

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	1,111,041 / 42,389	uid, perspective_idx, summary, doc_id, modality, url, domain (train), rank + text (eval; text inline for text docs)
`corpus_text` / `corpus_image`	train	631,377 / 380,672	doc_id, url, text (passage) / image (JPEG) — the paper's retrieval corpus: gold documents of every split plus distractors

The corpus is the 1.01M-document pool used for all results in the paper. Text documents are C4 passages (c4-<file>-<line>) shipped with their full text; images are shipped as JPEG re-encodings (longest side ≤ 512 px, quality 90, transparency composited on white) together with their source url. 12 images whose source returned HTML or gzip instead of an image were removed from the corpus and from gold/train (none were test gold).

from datasets import load_dataset
gold = load_dataset("seokwon99/Multi3IR", "gold", split="eval")
images = load_dataset("seokwon99/Multi3IR", "corpus_image", split="train")   # ~15 GB

Sources: Stack Exchange (CC BY-SA 4.0), C4 (ODC-BY), images © their owners, redistributed downscaled for research use only. 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}
}