You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ColPali Training Superset β€” 2M (metadata only)

Curated training data for ColPali-style late-interaction document retrievers. This repo is metadata-only β€” it does not ship the ~1.3 TB of image bytes. What's here is enough to rebuild the full superset bit-for-bit, plus pre-computed subset indices for fast lazy selection.

What you get

Layer Source Query rows
L1 6 curated sources (colpali, VisRAG x2, vdr-multilingual, tatdqa, tabfquad) ~655k
L2 Docmatix-IR (2 queries per PDF) with K=4 mined hard negatives sidecar ~2,164k
Superset ~2,820k queries (5,568,133 total rows)

L1 is structured by 18 K-means clusters (cluster_id 0-17, sizes 6.6k-40.5k). L2 docmatix queries carry cluster_id = -1 (never clustered).

Subsets shipped

The canonical sampling is v3 β€” cluster_uniform_capped_3-15pct (uniform-per-cluster with floor 3% / ceiling 15% per cluster, per Nemotron ColEmbed v2, arXiv:2602.03992 Β§3.4.3). Cluster -1 (docmatix) is capped at 15% of target so it doesn't dominate the training mix.

Subset Sampled queries Total rows Sidecar docmatix %
500k_v3 500,000 653,558 40,255 8.05%
750k_v3 572,496 β˜… 971,169 112,500 15% of target
1m_v3 609,996 β˜… 1,123,652 150,000 15% of target

β˜… 750k and 1m undershoot their nominal targets β€” the L1 pool (462k) is the structural ceiling under the 15% docmatix cap. See RECIPE.md for details. The undershoot is the correct output under the cap, not a bug.

Legacy v1 (proportional sampling) indices are not shipped in this repo (only the v3 capped-uniform indices are); the v1 algorithm is documented in RECIPE.md for forensic context.

Composition note (multilingual)

subset_500k_v3 is multilingual, not English-only: ~46.8% English, ~45% non-English (native vdr_multilingual/finance_commons + sibling-project translations of 5 EN sources), ~8% untagged Docmatix. It is v3, not the rejected experimental v4 (which would have translated every source). Full measured breakdown: subset_indices/subset_500k_v3/COMPOSITION.md.

Quickstart β€” lazy selection

from datasets import load_from_disk
from huggingface_hub import snapshot_download
import numpy as np
import pyarrow.parquet as pq

# 1. Pull this metadata repo (~150 MB)
repo_dir = snapshot_download(
    repo_id="athrael-soju/ColPali_Training_Superset_2m",
    repo_type="dataset",
)

# 2. Point to your locally rebuilt superset (1.3 TB, see RECIPE.md)
superset = load_from_disk("/your/path/datasets/superset_2m_clean/dataset")

# 3. Apply v3 indices for zero-copy subset views
indices = np.load(f"{repo_dir}/subset_indices/subset_500k_v3/indices.npy")
subset = superset.select(indices)
sidecar = pq.read_table(f"{repo_dir}/subset_indices/subset_500k_v3/sidecar.parquet")

Full rebuild (~7-8h)

See RECIPE.md. Phases: manifest -> docmatix-IR join -> assemble L1+L2 -> dedup -> derive subset indices. The deterministic scripts/subsample.py (seed=42) reproduces the v3 indices bit-for-bit.

License

Each source dataset retains its own license. The recipe + index files in this repo are Apache-2.0.

Citation

Please cite the original sources (Docmatix, Tevatron docmatix-ir, ColPali, VisRAG, VDR, TatDQA, TabFquad) and Nemotron ColEmbed v2 (arXiv:2602.03992) for the sampling recipe.

Downloads last month
589

Paper for athrael-soju/ColPali_Training_Superset_2m