A subset of LAION recaptioned with Gemma-3-27B-IT. Each image is paired with a single descriptive sentence (~20–30 words) generated by Gemma using the prompt:

Describe this image in 20-30 words. Do not include any preamble or introduction, just the description.

This dataset is used by the PuzzleBench project for vision/text alignment training.

Stats


Samples	508,025
Image resolution	up to 1500×1500 (varies)
Format	JPEG
Caption type	Gemma-3-27B recaption (declarative, ~130 chars avg)
Shards	30 webdataset `.tar` files (~2.4 GB each)
Total size	~72 GB

Format: WebDataset

The dataset ships as 30 webdataset-compatible tar shards, the de facto standard for large image-text corpora (used by LAION-400M, LAION-5B, CLIP training pipelines, etc.). Each shard is a flat tar containing:

shards-00000.tar
├── 000000.jpg     # raw image bytes
├── 000000.txt     # caption (utf-8)
├── 000001.jpg
├── 000001.txt
└── ...

A metadata.csv with the same image_id, url, caption is also shipped alongside for non-webdataset consumers.

Usage

With `webdataset` (streaming, no extraction needed)

import webdataset as wds

url = "https://huggingface.co/datasets/PuzzleBench/Recaptioned_LAION/resolve/main/shards-{00000..00029}.tar"
ds = (
    wds.WebDataset(url)
    .decode("pil")
    .to_tuple("jpg", "txt")
)

for image, caption in ds:
    # image: PIL.Image
    # caption: str
    ...

With `huggingface_hub` snapshot + local extraction

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="PuzzleBench/Recaptioned_LAION",
    repo_type="dataset",
    local_dir="recaptioned_laion",
)

With `pandas` (captions only, no images)

import pandas as pd
df = pd.read_csv("hf://datasets/PuzzleBench/Recaptioned_LAION/metadata.csv")
# columns: image_id, url, caption

License

This dataset of recaptions is released under CC BY 4.0. The underlying images are sourced from public LAION URLs and retain their original licensing — please consult the LAION usage terms when redistributing or training on the image content.

Citation

If you use this dataset, please cite the TDDN paper (as mentioned below), cite/credit the LAION-5B paper and corpus, and google/gemma-3-27b-it model.

Citation

@article{patnala2026tddn,
      title={{TDDN}: {T}ext-aligned {D}iffused {D}INO {N}etwork for Puzzle Understanding}, 
      author={Harsha Patnala and Debopriyo Banerjee and Ayush Sunil Munot and Somak Aditya},
      year={2026},
      journal={arXiv:2609.07937}
      eprint={2609.07937},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.07937}, 
}