lightonai/NanoViDoRe_v3

Nano ViDoRe Benchmark v3

A subsampled version of the ViDoRe Benchmark V3 for fast visual-document-retrieval evaluation during training. We keep a shared pool of 4,000 pages as the corpus so it can be encoded quickly while still being large enough to be a meaningful retrieval task. Distractor documents are mined from the original corpora to provide a challenging negative pool. The queries and qrels are sampled from the original datasets to preserve the benchmark's multilingual and multi-format characteristics.

Contents

Three configs, all with a single test split:

config	rows	description
corpus	4,000	page images + markdown text, merged across the 8 public V3 tasks
queries	400	50 per source dataset, uniformly distributed across languages (en/fr/de/es/it/pt)
qrels	2,061	graded relevance (1/2) from the original annotations

Provenance columns (source_dataset, source_corpus_id, source_query_id) map every row back to the original ViDoRe v3 datasets. is_judged marks gold pages vs. mined negatives. The markdown column makes the same pool usable as a text-retrieval benchmark.

How it was built

Queries. We sample 50 distinct queries per dataset, equally split across languages (never two translations of the same query), stratified by query_format to preserve the benchmark's question/instruction/keyword mix.

Document pool. All 1,782 pages judged as positives for the sampled queries are included in the pool. The remaining 2,218 pages are hard negatives mined with vultr/VultronRetrieverCore-Qwen3.5-4.5B: each query retrieves within its own source corpus, and we keep its top-200 retrieved documents (excluding positives) as distractor candidates. The 2,218 slots are then apportioned to each dataset in proportion to its original corpus size, from 604 for industrial (5,244 source pages) down to 128 for hr (1,110), so every dataset keeps roughly the same relative weight it has in the full benchmark. Within a dataset, negatives are selected iteratively over its 50 queries: each query in turn adds its next-best-ranked candidate that is not already pooled, until the total budget is filled. Each query therefore contributes an equal share of its hardest negatives (~3–12 pages, depending on the dataset).

Deduplication. On insert, a candidate negative is dropped if its normalized markdown matches an already-pooled page exactly (xxHash-64) or by 13-gram containment ≥ 50% in either direction, as described in our DenseOn-LateOn blog post and our paper. In total 218 near-duplicates were skipped.

Evaluation

Score each query against the full 4,000-page merged corpus; its qrels reference only pages from its own source dataset. Report nDCG@10 per source_dataset and the 8-task mean. Absolute scores are expected to be higher than the full benchmark because of the smaller corpus, so treat numbers as a training-time signal and confirm final results on full ViDoRe v3.

Loading

from datasets import load_dataset

corpus = load_dataset("lightonai/NanoViDoRe_v3", "corpus", split="test")
queries = load_dataset("lightonai/NanoViDoRe_v3", "queries", split="test")
qrels = load_dataset("lightonai/NanoViDoRe_v3", "qrels", split="test")

Source datasets

Check out the ViDoRe V3 collection for licenses and dataset details.

Citation

@misc{loison2026vidorev3,
  title         = {ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios},
  author        = {António Loison and Quentin Macé and Antoine Edy and Victor Xing and Tom Balough and Gabriel Moreira and Bo Liu and Manuel Faysse and Céline Hudelot and Gautier Viaud},
  year          = {2026},
  eprint        = {2601.08620},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2601.08620},
}