athrael-soju/vdr-train-500k

VDR Train 500k

VDR Train 500k has 499,704 search queries, each paired with the document page image that answers it. Queries are in English (73%), Spanish, German, French, Italian and Portuguese.

Construction

1. The source collections are merged into one pool of query-page pairs.
2. A vision encoder embeds every English page, PCA reduces the embeddings to 50 dimensions and K-Means groups them into 14 clusters. English pairs are sampled evenly across the clusters, with each cluster making up 3% to 15% of the sample. VDR-multilingual pairs are added in proportion to their languages.
3. Pairs whose query exactly matches a query in any of the 22 ViDoRe V1, V2 and V3 test repositories are removed.
4. Gemma 4 31B translates 22,985 English queries, sampled evenly across the clusters, into French, Spanish, German, Italian and Portuguese.

Columns

Column	Type	Content
query	string	Search query
positive_image	image	Page that answers the query
positive_text	string	Text of that page
lang	string	Query language: en, es, de, fr, it or pt
source_dataset	string	Source collection
original_id	string	Id of the pair in its source collection
cluster_id	int64	Page cluster (0-13) from step 2 of Construction
query_id	string	Query id

Hard negatives

hard_negatives.parquet gives each query up to four hard negatives: pages a late-interaction retriever ranked high for that query, though still below its positive page. The retriever found four for 495,585 queries, one to three for 1,269 and none for 2,850.

Column	Type	Content
query_row_idx	int64	Row of the query in the dataset
neg_row_idxs	list of int64	Rows whose positive_image is a negative page for the query
positive_score	float	Retriever score of the positive page
neg_scores	list of float	Retriever scores of the negatives, in the same order

Loading

dataset/ holds 337 Arrow shards (168 GB). Row numbers in hard_negatives.parquet follow the order in which load_from_disk returns the dataset.

import pandas as pd
from datasets import load_from_disk
from huggingface_hub import snapshot_download

local = snapshot_download("athrael-soju/vdr-train-500k", repo_type="dataset")
ds = load_from_disk(f"{local}/dataset")
negatives = pd.read_parquet(f"{local}/hard_negatives.parquet")

row = negatives.iloc[0]
query = ds[int(row.query_row_idx)]["query"]
negative_pages = [ds[int(i)]["positive_image"] for i in row.neg_row_idxs]

Sources

source_dataset	Source	Rows	License
docmatix_ir	Tevatron/docmatix-ir	149,538	MIT
visrag_synthetic	openbmb/VisRAG-Ret-Train-Synthetic-data	81,125	Source terms
visrag_in_domain	openbmb/VisRAG-Ret-Train-In-domain-data	62,758	Source terms
colpali_train_set	vidore/colpali_train_set	41,466	Original license of each subset
wiki_ss_nq	Tevatron/wiki-ss-nq	26,892	Source terms
vdr_multilingual_{en,es,de,fr,it}	llamaindex/vdr-multilingual-train	23,000	Apache-2.0
<source>_translated_{fr,es,de,it,pt}	Step 4 of Construction	114,925	Terms of the translated pair

License

Each pair keeps the license of its source dataset, listed under Sources. The hard negatives in hard_negatives.parquet and the cluster_id column are released under Apache-2.0.