PaS-B

Birds-domain dataset used to train the PaS-B checkpoint from the Precision at Scale paper (1.2M images).

Paper

• Paper (Pattern Recognition, 2026): https://doi.org/10.1016/j.patcog.2025.112236
• Official code: https://github.com/jesusmolrdv/Precision-at-Scale

@article{rodriguezdevera2026precision,
  title   = {Precision at scale: Domain-specific datasets on-demand},
  author  = {Rodr{\'i}guez-de-Vera, Jes{\'u}s M. and Estepa, Imanol G. and Saras{\'u}a, Ignacio and Nagarajan, Bhalaji and Radeva, Petia},
  journal = {Pattern Recognition},
  volume  = {171},
  pages   = {112236},
  year    = {2026},
  publisher = {Elsevier},
  doi     = {10.1016/j.patcog.2025.112236}
}

arXiv preprint

https://arxiv.org/abs/2407.03463

Composition

Published as two splits:

• web: 358,292 rows. URL-only, no image bytes hosted. Reconstruct locally with img2dataset. Obtained directly from Re-LAION-5B.
• synthetic: 856,467 rows. Actual image bytes, generated with stabilityai/stable-diffusion-2-1-base.

Data Owner Opt-Out

The web split contains only URLs pointing at third-party web images (no image bytes), sourced from Re-LAION-5B. If you own an image referenced by a URL in this split and want it removed:

• From the upstream index: submit a removal request via Spawning's haveibeentrained.com, the official opt-out mechanism LAION uses to build each Re-LAION-5B revision.
• From this specific redistribution: since this split is a snapshot taken at a point in time, an upstream opt-out registered afterward won't automatically propagate here. Open an issue on the official repository with the URL(s) or image_name(s) in question, and they will be removed from future releases of this dataset.

Usage

Load the synthetic split directly (image bytes are hosted):

from datasets import load_dataset

synthetic = load_dataset("jesusmolrdv/pas-b", "synthetic", split="train")

The web split has no hosted image bytes, only URLs. Reconstruct it locally with img2dataset:

from datasets import load_dataset

web = load_dataset("jesusmolrdv/pas-b", "web", split="train").to_pandas()
web[["url", "image_name"]].to_parquet("PaS-B-web-urls.parquet")

img2dataset --url_list PaS-B-web-urls.parquet --input_format parquet \
    --url_col url --save_additional_columns '["image_name"]' \
    --output_folder PaS-B_web_images --output_format webdataset