BootsofLagrangian/datacomp-recap-qwen3p5-35b-a3b

DataComp recaptions with Qwen3.5-35B-A3B

Dataset datacomp: 325.623 Million caption rows.

This public caption-only repository contains 325,623,447 generated caption rows for 325,472,073 normalized-URL identities. It unifies the selected existing recap pack, the later generated/recovery records, and the managed legacy v2 recap pack without exposing historical surface names. Distinct stripped captions for the same URL are retained; only exact normalized-URL plus stripped-caption duplicates are removed.

The matching source URL (sourceurlnorm) and stable URL identity (stableimageid and sourceurlsha1) are retained. No image payload, image locator, image hash, or claim of image licensing is included. Individual source-image rights remain with their respective rights holders. Obtain bytes independently from source URLs, for example with img2dataset, and review rights before any use. This caption-only release does not reconcile a caption against image availability, image content, or a source-universe membership table.

Matching image payload

The separately hosted gated image materialization is available on ModelScope: `BootsofLagrangian/datacomp-1b-url-dedup-webp90-min256px-noresize`. Reconcile the two releases by normalized source URL identity (source_url_sha1 / stable_image_id), not by Parquet row order. The image repository is an independently materialized preservation surface; its availability does not change the image-rights and preprocessing requirements above.

The `generation_config.yaml` file documents the Qwen3.5-35B-A3B-FP8 policy used for paper-aligned project recap runs. The legacy v2 import is accepted only when every source row has a nonblank recap_model provenance marker; this release does not infer an exact model revision or numeric precision for a legacy row beyond that marker.

This release is for research and audit reproducibility. It is not training-ready: users must perform semantic post-processing, safety and policy filtering, URL/content identity verification, and image-rights review before training use.

CC BY 4.0 applies only to the generated caption text in this repository. It does not license source images, URLs, or third-party metadata; rights in each source image remain with the respective rights holder.

@misc{oh2026matchedbudget,
  title={A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions},
  author={Giyeong Oh and Junghun Park and Yuhan Bae and Youngjae Yu},
  year={2026}
}