BootsofLagrangian/laion-highres-aesthetic-recap-qwen3p5-35b-a3b

LAION Highres Aesthetic recaptions with Qwen3.5-35B-A3B

Dataset laion-highres-aesthetic: 96.072 Million caption rows.

This public caption-only repository contains 96,071,504 generated captions for 82,690,560 image instances and no image payload. It includes 13,380,944 additional distinct caption variants. The matching gated image payload is hosted on ModelScope at `BootsofLagrangian/laion-highres-aesthetic-webp90-hybrid-resolution`. The Hugging Face image repository `BootsofLagrangian/laion-highres-aesthetic-webp90-noresize` is a pointer to that hosted payload. Rows match the image release through image_shard and image_member; URL and content hashes support independent reconciliation.

The image release preserves 124,416,249 distinct materialized image instances spanning 121,348,952 stable URL identities, including 3,067,297 additional instances that share a URL identity. A caption binds to its exact image instance when its recorded WebDataset locator matches the release; otherwise a URL-keyed caption falls back deterministically to the lexicographically smallest source key. The published caption rows contain 96,071,504 exact-locator bindings and 0 deterministic URL fallbacks. After applying that rule, 41,725,689 image instances have no direct caption row. This total can include both additional instances sharing a URL and stable URLs without a published caption, so an image-to-caption inner join must not be interpreted as full image coverage.

Original source metadata and our recap

The image repository preserves the source-provided caption field, when present, as source_caption. That field is original source metadata: we did not author, replace, or correct it. The caption_text field in this separate repository is our generated recap. Only that generated recap is an output of the method described below.

Load captions only

from datasets import load_dataset

captions = load_dataset(
    "BootsofLagrangian/laion-highres-aesthetic-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 ModelScope 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 instance.

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, image preprocessing, and reproduction hyperparameters.

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. It does not license source images or third-party metadata.

@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}
}