Hugging Face's logo

Datasets:
udgamlabs
/
chitra-tiny-data

Tasks:
Text-to-Image
Modalities:
Image
Languages:
English
Size:
1M<n<10M
ArXiv:
Tags:
udgamlabs
image
webdataset
public-domain
cc0
latents
Libraries:
WebDataset
License:
Dataset card Files
xet
 Community
chitra-tiny-data / README.md
jiteshkprasad's picture
jiteshkprasad
Update dataset card: tested examples
8c6154d verified
preview code
|
Raw History Blame Contribute Delete
11.6 kB
metadata 
pretty_name: Chitra Tiny training set
license: other
license_name: cc0-pdm-cdla-permissive-2.0-mit
license_link: >-
  https://huggingface.co/datasets/udgamlabs/chitra-tiny-data/blob/main/LICENSES/README.md
language:
  - en
task_categories:
  - text-to-image
size_categories:
  - 1M<n<10M
tags:
  - udgamlabs
  - image
  - webdataset
  - public-domain
  - cc0
  - latents
configs:
  - config_name: default
    data_files:
      - split: train
        path: train/*.tar
      - split: validation
        path: validation/*.tar

Chitra Tiny training set · Udgam Labs

UDGAM LABS · Image-text dataset · 4.7 million images

Chitra Tiny training set

The images that Chitra Tiny, our 212M-parameter text-to-image model, was trained on: 4,691,814 pictures at 256×256, each with its caption and its SD-VAE latent, plus a held-out split of 14,384 images that the model never saw. Every image was dedicated under CC0 or marked public domain (PDM) by the source. The images come from two public collections curated by Spawning, PD12M and Megalith-CC0. We resized and encoded them, and before release we scanned every image and removed 29,460 that failed a safety or licence check.

See the model: Chitra Tiny on Hugging Face · in your browser · code on GitHub · udgamlabs.com/models/chitra-tiny

Quick start

Stream a few samples without downloading the whole set (159 GB):

import numpy as np
from datasets import load_dataset

ds = load_dataset("udgamlabs/chitra-tiny-data", split="train", streaming=True)
for s in ds.take(3):
    image, caption = s["jpg"], s["txt"]                       # PIL image (256×256), str
    latent = np.asarray(s["npy"], dtype=np.float16)           # (4, 32, 32)
    print(s["__key__"], s["json"]["license"], image.size, latent.shape, caption[:60])

Or read the tar shards directly with webdataset:

import webdataset as wds

url = "https://huggingface.co/datasets/udgamlabs/chitra-tiny-data/resolve/main/train/train-{00000..00143}.tar"
ds = wds.WebDataset(url, shardshuffle=True).shuffle(5000).decode("pil").to_tuple("jpg", "txt", "npy", "json")

Decode a latent back to pixels with the same VAE:

import torch
from diffusers import AutoencoderKL

vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse")
x = vae.decode(torch.from_numpy(latent).float()[None] / 0.18215).sample   # (1, 3, 256, 256), values in [-1, 1]

What is in it

Split Images From PD12M From Megalith-CC0 Shards
train 4,691,814 2,491,765 2,200,049 144 (train/train-00000.tar … train-00143.tar)
validation 14,384 7,476 6,908 2 (validation/validation-00000.tar, -00001.tar)

Each sample has four files with the same key (the source dataset's own image id):

File What it is
KEY.jpg The image: 256×256 RGB, short side resized to 256 and centre-cropped, JPEG quality 90
KEY.txt The caption, unchanged from the source
KEY.npy The image's latent from stabilityai/sd-vae-ft-mse: the posterior mean × 0.18215, float16, shape 4×32×32
KEY.json source_dataset (PD12M or Megalith-CC0) and license (CC0-1.0 or PDM-1.0)

Licence index. index/train.parquet and index/validation.parquet have one row per image: key, split, shard, source_dataset, license, license_url, image_url (the source's copy of the original), source_collection (for PD12M, the institution or site it came from), original_width, original_height and source_hash. Load it with load_dataset("parquet", data_files="hf://datasets/udgamlabs/chitra-tiny-data/index/train.parquet") or with pandas.

Licences: 3,706,589 images (78.8%) are CC0 and 999,609 (21.2%) carry the Public Domain Mark. Within PD12M the largest sources are Wikimedia Commons (56.5%), iNaturalist (20.5%), Rijksmuseum (1.3%).

Where it falls short

  • Narrow in what it shows. PD12M is mostly Wikimedia Commons, iNaturalist and museum and archive collections, so it leans towards historic, Western and museum imagery, plants and animals. Megalith-CC0 is everyday Flickr photographs, many of them of people.
  • Small and cropped. 256 px centre crops lose the edges of wide and tall images and all fine detail.
  • An older latent space. The latents are for SD-VAE ft-mse (8× downsampling, 4 channels). Most newer image models use other autoencoders, so you may need to re-encode the images.
  • Machine-written captions. PD12M's captions were written by Florence-2-large. The source does not say which model wrote Megalith-CC0's captions; they are much longer (in the validation split, a median of 89 words against 37 for PD12M) and often start "The image shows". Both can be wrong about what is in a picture.
  • The safety scan is automatic and imperfect. Two classifiers scored every image. They rarely agree on which images to flag (2,657 images were flagged by both, 26,946 by either), which suggests both miss some unsafe images and flag some harmless ones. We removed anything either flagged. No image was checked against lists of known abuse images (hash matching); report anything you find (see Takedown below).
  • Licence status is the source's label, not a guarantee. PD12M's authors note they cannot guarantee the licence of every work. The Public Domain Mark records the source's view that a work's copyright has expired, and that can depend on the country.
  • A frozen copy. The sources remove items they learn are infringing or unsafe. Those removals do not reach this copy by themselves; we apply them when we hear of them.

How it was made

Sources. PD12M (Spawning): metadata files 000–024, the first 2.5 million of its 12.4 million rows, for training, and file 123 for the held-out split. Megalith-CC0 (Spawning, a CC0-only selection of Megalith-10m): all nine parts, with the first 10,000 rows of part 8 held out. Images were downloaded from the sources' own hosted copies.

Selection. Kept rows whose licence is CC0 or the Public Domain Mark and whose caption has at least 8 words. Images that failed to download were skipped (about 6% of Megalith-CC0, evenly across its files). Chitra Tiny was trained on the resulting 4,721,166 images: 52.9% from PD12M and 47.1% from Megalith-CC0.

Processing. Each image was decoded, resized so its short side is 256 px (bicubic), centre-cropped to 256×256 and saved as JPEG quality 90; the latent is the SD-VAE ft-mse posterior mean, scaled by 0.18215. Apart from the clean-up below, this is the data Chitra Tiny was trained and evaluated on. Its FID numbers use 5,000 images from the validation split; the clean-up took out 108 validation images, so a few of those may be missing here.

Release clean-up and safety scan (October 2026). Before publishing we went over all 4,735,658 images once:

Step What it did Images removed
Hidden text in image files Removed JPEG comment and metadata segments from 246,489 files at the byte level; the pixels are unchanged (checked for each file) —
Copyright notice Removed images whose embedded comment carried a copyright notice, which conflicts with their CC0 or PDM label 865
Safety classifiers Scored every image with Falconsai/nsfw_image_detection and AdamCodd/vit-base-nsfw-detector (both Apache-2.0); removed every image either scored above 0.5 26,946
Caption rule Removed every image whose caption pairs a word for a child with a word for nudity, underwear or swimwear. No image was viewed; the rule is applied to the text only 1,347
Personal data in captions Removed images whose caption contains an email address, a phone number or a key-like string 567
Licence check Matched every image to the source's metadata; removed any whose id or licence did not check out 0
Damaged samples Removed undecodable images or unreadable latents 0
Total Some images failed more than one check 29,460 (0.6%)

Intended use and limitations

For training and studying small text-to-image and latent-diffusion models, for teaching, and for reproducing Chitra Tiny. Not for identifying or profiling the people who appear in the photos, and not as a source of facts about what an image shows. If you build a public product on it, add your own safety filtering.

Takedown

If an image or caption should not be here (an unsafe image, a person who wants their photo removed, a work that is not in the public domain), open a discussion on this dataset's Community tab or email jitesh@udgamlabs.com. We take reported items out of this copy, and also items that PD12M or Megalith-CC0 take out of theirs.

Licence

Each part has its own terms; see LICENSES/README.md. In short:

Part Terms
Images Dedicated under CC0 1.0 or marked public domain (PDM 1.0) by the source; per image in the index
PD12M captions and metadata CDLA-Permissive-2.0
Megalith-CC0 captions and metadata MIT
What Udgam Labs adds (latents, index, per-sample .json, this card) CDLA-Permissive-2.0

CC0 and the Public Domain Mark cover copyright only. They do not cover the privacy or publicity rights of people shown in the images, or trademarks, logos and insignia.

Citation 

@misc{chitratinydata2026,
  title  = {Chitra Tiny training set: 4.7M public-domain and CC0 images at 256 px with captions and SD-VAE latents},
  author = {Udgam Labs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/udgamlabs/chitra-tiny-data}
}

Please also cite the sources:

@misc{meyer2024publicdomain12mhighly,
  title         = {Public Domain 12M: A Highly Aesthetic Image-Text Dataset with Novel Governance Mechanisms},
  author        = {Jordan Meyer and Nick Padgett and Cullen Miller and Laura Exline},
  year          = {2024},
  eprint        = {2410.23144},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2410.23144}
}

Megalith-CC0: https://huggingface.co/datasets/Spawning/megalith-cc0, derived from Megalith-10m by madebyollin (https://huggingface.co/datasets/madebyollin/megalith-10m).

Thanks to Spawning for PD12M and Megalith-CC0, to madebyollin for Megalith-10m, to Microsoft for Florence-2, to Stability AI for the SD-VAE, and to the authors of the two safety classifiers.

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com