Research

Datasets & code.

What the lab has released so far, and a working library of the datasets our members actually use — organised by area, with links to the source.

01 — From the lab

One release so far.

We are not going to overstate this. CAAI has not released large-scale datasets — we are a small lab and most of our output to date has been papers rather than corpora. One dataset is public, and more will follow from the projects currently running.

Dataset · Zenodo · July 2025

Synthetic Image Dataset of Five Object Classes Generated Using Stable Diffusion XL

A synthetic image corpus across five object classes, generated with Stable Diffusion XL and released openly for use in classification, augmentation and synthetic-data experiments.

Gurpreet Singh DOI 10.5281/zenodo.16414387

In preparation

  • Aligned image–music pairs. Being assembled for the image-to-music project, together with the listening-study protocol used to evaluate it.
  • Script-to-panel annotations. Narrative beats and panel layouts from the text-to-comic project, including character-consistency labels.

Neither is public yet. When they are, they will be released with a DOI, a licence and a datasheet describing how they were built and where they should not be used.

02 — Dataset library

The standard corpora, by area.

These are the datasets a new member is most likely to be pointed at in their first month. Licences vary — several require registration or an agreement before download, and a few are restricted to credentialed researchers. Check the terms on each source before you use one in a paper.

Computer Vision

Classification · Detection · Segmentation
ImageNet
Stanford / Princeton
14M+ labelled images across 20k categories. The benchmark that started modern deep learning; ILSVRC-2012 remains the standard pretraining set.
Download ↗
MS COCO
Microsoft
330k images with object segmentation, keypoints and five captions each — the default for detection, segmentation and captioning at once.
Download ↗
Open Images V7
Google
9M images with box, segmentation, relationship and localised narrative annotations. The largest openly licensed detection set.
Download ↗
ADE20K
MIT CSAIL
Scene parsing with dense pixel-level annotation over 150 semantic categories. The reference set for semantic segmentation.
Download ↗
Cityscapes
Daimler / TU Darmstadt
Urban street scenes from 50 cities with fine pixel annotation. Standard for autonomous-driving perception. Registration required.
Register ↗
CIFAR-10 / CIFAR-100
University of Toronto
60k 32×32 images in 10 or 100 classes. Small enough to iterate on a laptop — where most members run their first training loop.
Download ↗
MNIST
LeCun et al. · Hugging Face mirror
70k handwritten digits. Solved long ago, still the fastest way to check that a pipeline works end to end before you scale it.
Download ↗
Fashion-MNIST
Zalando Research
A drop-in MNIST replacement with clothing categories — same format, meaningfully harder.
Download ↗
CelebA
CUHK MMLab
200k celebrity faces with 40 attribute labels. Widely used for generative work — and worth reading the fairness critiques before you do.
Download ↗

04 — Working stack

What members actually run.

PyTorch

The lab's default framework. Nearly every project here is written against it.

pytorch.org ↗

TensorFlow

Used where an existing implementation or pretrained checkpoint calls for it.

tensorflow.org ↗

Transformers

Pretrained language, vision and multimodal models with a consistent interface.

huggingface.co ↗

Diffusers

Diffusion pipelines for image and audio generation — the basis of our synthetic data work.

huggingface.co ↗

Datasets

Loading, streaming and preprocessing corpora too large to hold in memory.

huggingface.co ↗

scikit-learn

Baselines, classical models and evaluation metrics. Always run the baseline first.

scikit-learn.org ↗

Weights & Biases

Experiment tracking, so that a result from three weeks ago can still be reproduced.

wandb.ai ↗

Google Colab

Free GPU access for members without local hardware. Where most first experiments run.

colab.research ↗

arXiv

Where the papers you are implementing live, and where preprints go first.

arxiv.org ↗

05 — Before you download

Three things that get papers rejected.

Licensing

Check the terms first.

"Publicly available" is not the same as "free to use however you like". Several datasets above permit academic research only, some require a signed agreement, and a few forbid redistribution of derived data entirely. Read the licence before the first experiment, not before submission.

Citation

Cite the dataset paper.

Every dataset here has an accompanying publication, and its authors are counting on that citation the same way you will count on yours. Cite the corpus, the version and the split you used — vague provenance is a reviewer's first complaint.

Ethics

Know who is in your data.

Web-scraped corpora contain people who never agreed to be there, and their skew becomes your model's behaviour. Report what your data covers and what it misses. This is not a formality here — it is one of our four research areas.

Releasing

Publish yours properly.

If a project produces a dataset, it gets a DOI, an explicit licence and a datasheet covering collection method, intended use and known limitations. That is the standard the lab holds its own releases to.

Missing something?

Tell us what should be on this list.

If you work with a corpus in your field that belongs here — especially outside vision and English-language NLP — send it to us and we will add it. The list is meant to be useful, not complete.