BootsofLagrangian/laion-aesthetics-recap-qwen3p5-35b-a3b

LAION-Aesthetics recaptions with Qwen3.5-35B-A3B

Dataset laion-aesthetics: 24.290 Million caption rows.

This public caption-only repository contains 24,290,381 generated captions for 23,687,875 image assets and no image payload. It includes 602,506 additional distinct caption variants. Rows match BootsofLagrangian/laion-aesthetics-webp90-min256px-noresize through image_shard and image_member; URL and content hashes support independent reconciliation.

Captions were produced with the same caption-generation method used in A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions: Qwen/Qwen3.5-35B-A3B-FP8, one long-caption stage, a 512-token output limit, Qwen3.5 release sampling defaults, free-form output, and vLLM 0.17.1 serving. Thinking was disabled. Output received whitespace stripping only; no semantic post-processing was applied. See generation_config.yaml for the system prompt, user prompt, and reproduction hyperparameters.

Original caption and our recap

The original caption provided by the LAION-Aesthetics source is the web-derived TEXT associated with each URL; it is preserved as source_caption in the paired image metadata. Our recap is the separate caption_text generated from the image with Qwen3.5 in a consistent, direct, prompt-like visual register. Our recap does not modify, replace, or claim to correct the original caption. URL identity and byte hashes let users audit the original and our recap separately.

Load captions only

from datasets import load_dataset

captions = load_dataset(
    "BootsofLagrangian/laion-aesthetics-recap-qwen3p5-35b-a3b",
    split="train",
    streaming=True,
)
row = next(iter(captions))
print(row["image_shard"], row["image_member"], row["caption_text"])

Images live in the separate gated repository. Join on image_shard + image_member; URL and content hashes provide independent reconciliation checks. More than one caption row may intentionally refer to the same image asset.

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.

Personal and sensitive information

Captions are released as generated. The captioner describes what an image shows and transcribes text that is clearly legible in it, so a caption can carry a name, a contact detail, an identifier, or other personal or sensitive content that appears in the source image. The caption text is kept unmodified so that each user applies the filter that fits their use; filtering for personal and sensitive content belongs to the post-processing required before any training use.

To request removal of a row, open a discussion in the Community tab of this repository and give the row keys (caption_record_id and source_url_sha1).

License

Generated caption text: CC-BY-4.0. Code and configuration files in this repository (for example generation_config.yaml): Apache-2.0. Source images and original captions keep the licenses of their source releases.

@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}
}