BootsofLagrangian/megalith-cc0-recap-qwen3p5-35b-a3b

Megalith CC0 recaptions with Qwen3.5-35B-A3B

Dataset megalith-cc0: 8.073 Million caption rows.

This caption-only repository contains 8,072,536 generated captions for 8,072,536 Megalith crop identities and no duplicated image payload. Images and original source captions remain in the pinned Spawning/pd-extended release at revision a5f67f80efc42a4951f65b304aa630f4f9fb15d7. image_shard and image_member identify the pinned source Parquet and row; asset_instance_id is the Megalith crop UUID. The upstream caption field is original source metadata, while caption_text here is our generated recap.

from datasets import load_dataset

captions = load_dataset(
    "BootsofLagrangian/megalith-cc0-recap-qwen3p5-35b-a3b", split="train", streaming=True
)
row = next(iter(captions))
print(row["asset_instance_id"], row["caption_text"])

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 leading/trailing 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 in this repository. It does not alter or relicense upstream 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}
}