RuDoRe: Russian Document Retrieval

RuDoRe is an evaluation dataset for retrieving document pages from Russian text queries. The corpus contains page images from presentations, reports, course materials, and other documents with tables, charts, diagrams, and complex layouts. A retrieval system ranks these pages by relevance to a query.

This repository contains the test split: 18,805 queries and 6,383 pages from 174 documents. The training corpus is not distributed because of restrictions on redistributing the source materials. The adapted models are rudore/colpali-v1.3 and rudore/colqwen3-2b.

Loading the dataset

Install the Hugging Face Datasets library in your project:

uv add datasets

from datasets import load_dataset

corpus = load_dataset("rudore/RuDoRe", "corpus", split="test")
queries = load_dataset("rudore/RuDoRe", "queries", split="test")

print(corpus.features)
print(queries.features)

Configuration	Split	Contents
corpus	test	One row per page, with a PNG image and page/document metadata.
queries	test	Query text and gold positive_passages references into the corpus.

Keep corpus ordering and passage references consistent when exporting or indexing the data. Rank pages against the full evaluation corpus rather than a separately sampled set of negatives.

Dataset size

Quantity	Test split
Documents	174
Page images	6,383
Query-page pairs	18,805
Topics	13

The separate training corpus contains 188,192 query-page pairs and 96,607 pages. Splitting is performed at the document level, so pages from the same source document remain in one partition.

Topic distribution

Per-topic sampling limits the contribution of large topics. Documents may have multiple topic labels, and the final counts are not uniform.

Topic	Test pairs
Academic course decks	3,589
Computer science	2,330
Natural sciences	2,075
Medicine and biology	1,523
Economics and business	1,382
History and geography	1,153
Architecture and urbanism	1,119
Professional workshops	1,104
Design and UX/UI	1,087
Psychology	1,039
Real estate marketing	1,003
BigTech reports	1,001
Startup pitch decks	400

Data collection and annotation

1. Collect documents. A plan covering 13 topics and 656 seed queries drives PDF, PPTX, and DOCX discovery through Yandex Search API. Content-hash deduplication and file validation retain 2,355 documents with 228,902 pages before downstream filtering.
2. Generate and validate queries. Qwen3-VL-32B-Instruct generates Russian queries from rendered page images. A separate judge prompt checks whether each query is relevant to its page and sufficiently specific. Accepted pairs receive the aligned label.
3. Split by document. A topic-aware partition keeps every page of a document in the same split and retains coverage of all 13 topics.
4. Rank and filter pairs. A frozen visual retriever ranks each target page within its split. The gold_in_top_100 filter retains pairs whose target is among the first 100 results. Text-only pages are removed from the supervision set; the construction pipeline also mines training negatives.
5. Sample by topic. Per-topic quotas reduce the dominance of large crawls and produce the final training and test supervision sets.
6. Rephrase queries. Queries are rewritten while preserving their information need, reducing reliance on wording copied from the source page. The main RuDoRe evaluation uses these rephrased queries. Original queries, called RuDoRe NR, are a separate diagnostic control and are not included here.

The source documents are real documents; the query-page supervision is generated automatically. Rephrasing changes query wording but does not establish that the queries represent real user traffic.

Human annotation audit

Two annotators independently reviewed a random sample of 150 judge-accepted pairs from the training data.

Measure	Result
Mean share judged relevant	89%
Raw agreement between annotators	94%

This audit measures relevance among accepted training pairs. It is not a direct estimate of test-set precision or of the judge's recall over rejected pairs.

Reference retrieval results

The primary metric is nDCG@5, where higher values indicate better ranking of relevant pages. The following results use the main RuDoRe evaluation with rephrased queries.

Model	nDCG@5
ColPali v1.3, base	0.5071
rudore/colpali-v1.3	0.7091
SauerkrautLM-ColQwen3-2b-v0.1, base	0.8394
rudore/colqwen3-2b	0.8979

Adapted checkpoints were selected by nDCG@5 on a held-out RuDoRe validation shard before evaluation on the final test set. Text-model comparisons in the associated experiments were conducted on X5-private, a separate held-out business-document corpus.

Additional base-model comparison

The following reported RuDoRe results come from a separate model-size comparison. All four checkpoints are base retrievers without RuDoRe LoRA adaptation; the metric is per-topic macro nDCG@5 on rephrased queries.

Base model	Model size	RuDoRe nDCG@5
SauerkrautLM-ColLFM2-450M-v0.1	450M	0.2029
SauerkrautLM-ColQwen3-2b-v0.1	2B	0.8645
SauerkrautLM-ColQwen3-4b-v0.1	4B	0.8859
SauerkrautLM-ColQwen3-8b-v0.1	8B	0.9197

This comparison reports 0.8645 for base ColQwen3-2B, while the reference evaluation above reports 0.8394. These values originate from different evaluation runs. The cause of the discrepancy remains unresolved, so the two result sets are kept separate rather than combined into a single ranking. The additional scores are reported from the recorded experiment results and have not been rerun for this card update.

Intended use and limitations

RuDoRe supports evaluation of Russian page-level visual retrieval and analysis of adaptation, document layout, and query wording. Keep this test split held out from training when reporting comparable results.

• The crawl reflects the selected topics and documents surfaced by Yandex Search. It emphasizes visually rich documents rather than all Russian-language search tasks.
• Automatically generated queries and judge labels can contain errors. The human audit covers a small sample of accepted training pairs.
• Filtering with a pretrained retriever can exclude valid queries that this retriever finds difficult, favoring examples it can already retrieve.
• Rephrasing reduces lexical copying; it does not guarantee removal of all wording overlap or generation artifacts.
• Near-duplicate pages, small text, and complex tables can make the exact relevant page difficult to identify.

License and source materials

The repository declares Creative Commons Attribution 4.0. Page images originate from third-party documents; retain their source attribution and check the applicable source-material terms before redistribution. Rights holders can request removal of specific pages through the repository's Community tab.

Authors

Roman Makeev and Egor Machnev.