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

Danbooru recaptions with Qwen3.5-35B-A3B

Dataset danbooru: 13.642 Million caption rows.

This public caption-only repository contains 13,642,389 generated captions for 11,263,131 Danbooru post identities and no image payload. It unifies the March corpus, the later June crawl, and the recovered cosplay-focused recap increment without exposing internal forward/backward/retry surfaces. Distinct captions for the same post are preserved as variants; only exact post+caption duplicates are removed.

image_shard and image_member bind 13,114,274 caption rows (96.129%) to the separate gated BootsofLagrangian/danbooru-202603-webp90-noresize image release. The remaining 528,115 caption rows (3.871%) retain source URLs but refer to source assets outside that frozen 202603 image materialization. Of 13,642,412 scanned records, 23 (0.000169%) were excluded for an empty caption or invalid post ID, and 0 exact asset-caption duplicates were removed.

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 tag-grounded 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 prompts, grounding context, image preprocessing, and reproduction hyperparameters.

Original tags and our recap

The original supervision provided by Danbooru is its structured tag metadata rather than a prose caption; those tags remain original source metadata. Our recap is the separate caption_text generated from the image with Qwen3.5. It uses selected original tags as grounding anchors while expressing the visible scene as continuous, prompt-like anime prose. Our recap does not modify or replace the original tags. General, character, copyright, and rating context can guide generation; artist tags are not supplied.

Character naming is conditional and best effort. The preprocessing algorithm supplies confirmed_characters only when the general tags contain solo. It normalizes and deduplicates base character names, excludes a character root marked as a cosplay target, and separately retains raw qualified character tags for costume or form context. The prompt asks the VLM to use a confirmed base name and consult the raw tags when selecting visibly matching variant detail, but output is not programmatically forced or repaired. Rows without qualifying metadata receive no confirmed name, and even eligible generations can omit or phrase a name differently. This release does not claim that every eligible caption contains a character name or that any included name is visually correct.

Load captions only

from datasets import load_dataset

captions = load_dataset(
    "BootsofLagrangian/danbooru-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. A non-null image_shard + image_member pair binds to that frozen image release. Rows with null image locators remain caption/provenance records for source assets outside the image freeze.

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 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}
}