The STL-10 dataset is an image recognition dataset for developing unsupervised feature learning, deep learning, self-taught learning algorithms. It is inspired by the CIFAR-10 dataset but with some modifications. In particular, each class has fewer labeled training examples than in CIFAR-10, but a very large set of unlabeled examples is provided to learn image models prior to supervised training. The primary challenge is to make use of the unlabeled data (which comes from a similar but different distribution from the labeled data) to build a useful prior. All images were acquired from labeled examples on ImageNet.

To use this dataset:

import tensorflow_datasets as tfds

ds = tfds.load('stl10', split='train')
for ex in ds.take(4):
 print(ex)

See the guide for more informations on tensorflow_datasets.

https://storage.googleapis.com/tfds-data/visualization/fig/stl10-1.0.0.png" alt="Visualization" width="500px">

This public dataset contains labels for the unlabeled 100,000 pictures in the STL-10 dataset.

The dataset is human labeled with AI aid through Etiqueta, the one and only gamified mobile data labeling application. stl10.py is a python script written by Martin Tutek to download the complete STL10 dataset. labels.json contains labels for the 100,000 previously unlabeled images in the STL10 dataset legend.json is a mapping of the labels used. stats.ipynb presents a few statistics regarding the 100,000 newly labeled images.

If you use this dataset in your research please cite the following:

@techreport{yagli2025etiqueta,
 author = {Semih Yagli},
 title = {Etiqueta: AI-Aided, Gamified Data Labeling to Label and Segment Data},
 year = {2025},
 number = {TR-2025-0001},
 address = {NJ, USA},
 month = Apr.,
 url = {https://www.aidatalabel.com/technical_reports/aidatalabel_tr_2025_0001.pdf},
 institution = {AI Data Label},
}

@inproceedings{coates2011analysis,
  title = {An analysis of single-layer networks in unsupervised feature learning},
  author = {Coates, Adam and Ng, Andrew and Lee, Honglak},
  booktitle = {Proceedings of the fourteenth international conference on artificial intelligence and statistics},
  pages = {215--223},
  year = {2011},
  organization = {JMLR Workshop and Conference Proceedings}
}

Note: The dataset is imported to Kaggle from: https://github.com/semihyagli/STL10-Labeled See also: https://github.com/semihyagli/STL10_Segmentation

If you have comments and questions about Etiqueta or about this dataset, please reach us out at contact@aidatalabel.com

![]() The STL-10 dataset is an image recognition dataset for developing unsupervised feature learning, deep learning, self-taught learning algorithms. It is inspired by the CIFAR-10 dataset but with some modifications. In particular, each class has fewer labeled training examples than in CIFAR-10, but a very large set of unlabeled examples is provided to learn image models prior to supervised training. The primary challenge is to make use of the unlabeled data (which comes from a similar but different distribution from the labeled data) to build a useful prior. We also expect that the higher resolution of this dataset (96x96) will make it a challenging benchmark for developing more scalable unsupervised learning methods. Overview 10 classes: airplane, bird, car, cat, deer, dog, horse, monkey, ship, truck. Images are 96x96 pixels, color. 500 training images (10 pre-defined folds), 800 test images per class. 100000 unlabeled images for uns

The dataset used in the paper is CIFAR-10 and STL-10, which are commonly used datasets for image classification tasks.

The dataset used in the paper is CIFAR-10, STL-10, and ImageNet.

STL10-Labeled

This public repo contains labels for the unlabeled pictures in the stl10 dataset. Please refer to files Files and versions tab above. You can also refer to my original repo https://github.com/semihyagli/STL10-Labeled Please consider sponsoring this repo so that we can continue to develop high-quality datasets for the ML/AI research. To become a sponsor: GitHub Sponsors Buy me a coffee You can also sponsor us by downloading our free application, Etiqueta, to your… See the full description on the dataset page: https://huggingface.co/datasets/semihyagli/STL10-Labeled.

The dataset used in the paper is MNIST, CIFAR10 and STL10. These are datasets for image classification tasks.

The dataset used in this paper is a collection of images from the STL-10 dataset, preprocessed and used for training and evaluation of the proposed diffusion spectral entropy and diffusion spectral mutual information methods.

Dataset Card for STL-10 Cleaned (Deduplicated Training Set)

Paper | Code

  Dataset Description

This dataset is a modified version of the STL-10 dataset. The primary modification involves deduplicating the training set by removing any images that are exact byte-for-byte matches (based on SHA256 hash) with images present in the original STL-10 test set. The dataset comprises this cleaned training set and the original, unmodified STL-10 test set. The goal is to provide a… See the full description on the dataset page: https://huggingface.co/datasets/Shu1L0n9/CleanSTL-10.

STL10 - Segmentation

Please consider sponsoring this repo so that we can continue to develop high-quality datasets for the AI and ML research. To become a sponsor: GitHub Sponsors Buy me a coffee You can also sponsor us by downloading our free application, Etiqueta, to your devices: Etiqueta on iOS or Apple Chip Macs Etiqueta on Android This repo contains segmented images for the labeled part of the STL-10 Dataset.
If you are looking for STL10-Labeled variant of the dataset… See the full description on the dataset page: https://huggingface.co/datasets/semihyagli/STL10-Segmented.

The MNIST, KMNIST, FashionMNIST, STL-10 and CIFAR-10 datasets are used for few-shot learning experiments.

Abstract

In the last years, neural networks have evolved from laboratory environments to the state-of-the-art for many real-world problems. Our hypothesis is that neural network models (i.e., their weights and biases) evolve on unique, smooth trajectories in weight space during training. Following, a population of such neural network models (refereed to as “model zoo”) would form topological structures in weight space. We think that the geometry, curvature and smoothness of these structures contain information about the state of training and can be reveal latent properties of individual models. With such zoos, one could investigate novel approaches for (i) model analysis, (ii) discover unknown learning dynamics, (iii) learn rich representations of such populations, or (iv) exploit the model zoos for generative modelling of neural network weights and biases. Unfortunately, the lack of standardized model zoos and available benchmarks significantly increases the friction for further research about populations of neural networks. With this work, we publish a novel dataset of model zoos containing systematically generated and diverse populations of neural network models for further research. In total the proposed model zoo dataset is based on six image datasets, consist of 24 model zoos with varying hyperparameter combinations are generated and includes 47’360 unique neural network models resulting in over 2’415’360 collected model states. Additionally, to the model zoo data we provide an in-depth analysis of the zoos and provide benchmarks for multiple downstream tasks as mentioned before.

Dataset

This dataset is part of a larger collection of model zoos and contains the zoos trained on the labelled samples from STL10. All zoos with extensive information and code can be found at www.modelzoos.cc.

This repository contains the raw model zoos as collections of models (file names beginning with "cifar_"). Zoos are trained with small and large CNN models, in three configurations varying the seed only (seed), varying hyperparameters with fixed seeds (hyp_fix) or varying hyperparameters with random seeds (hyp_rand). Due to the large filesize, the preprocessed datasets are hosted in a separate repository. The index_dict.json files contain information on how to read the vectorized models.

For more information on the zoos and code to access and use the zoos, please see www.modelzoos.cc.

We introduce PiCCL (Primary Component Contrastive Learning), a self-supervised contrastive learning framework that utilizes a multiplex Siamese network structure consisting of many identical branches rather than 2 to maximize learning efficiency. PiCCL is simple and light weight, it does not use asymmetric networks, intricate pretext tasks, hard to compute loss functions or multimodal data, which are common for multiview contrastive learning frameworks and could hinder performance, simplicity, generalizability and explainability. PiCCL obtains multiple positive samples by applying the same image augmentation paradigm to the same image numerous times, the network loss is calculated using a custom designed Loss function named PiCLoss (Primary Component Loss) to take advantage of PiCCL’s unique structure while keeping it computationally lightweight. To demonstrate its strength, we benchmarked PiCCL against various state-of-the-art self-supervised algorithms on multiple datasets including CIFAR-10, CIFAR-100, and STL-10. PiCCL achieved top performance in most of our tests, with top-1 accuracy of 94%, 72%, and 97% for the 3 datasets respectively. But where PiCCL excels is in the small batch learning scenarios. When testing on STL-10 using a batch size of 8, PiCCL still achieved 93% accuracy, outperforming the competition by about 3 percentage points.

QTL Report

The dataset used in the paper is not explicitly described, but it is mentioned that the authors used CIFAR-10, CIFAR-100, and STL-10 datasets for training and testing the embedding functions.

We introduce PiCCL (Primary Component Contrastive Learning), a self-supervised contrastive learning framework that utilizes a multiplex Siamese network structure consisting of many identical branches rather than 2 to maximize learning efficiency. PiCCL is simple and light weight, it does not use asymmetric networks, intricate pretext tasks, hard to compute loss functions or multimodal data, which are common for multiview contrastive learning frameworks and could hinder performance, simplicity, generalizability and explainability. PiCCL obtains multiple positive samples by applying the same image augmentation paradigm to the same image numerous times, the network loss is calculated using a custom designed Loss function named PiCLoss (Primary Component Loss) to take advantage of PiCCL’s unique structure while keeping it computationally lightweight. To demonstrate its strength, we benchmarked PiCCL against various state-of-the-art self-supervised algorithms on multiple datasets including CIFAR-10, CIFAR-100, and STL-10. PiCCL achieved top performance in most of our tests, with top-1 accuracy of 94%, 72%, and 97% for the 3 datasets respectively. But where PiCCL excels is in the small batch learning scenarios. When testing on STL-10 using a batch size of 8, PiCCL still achieved 93% accuracy, outperforming the competition by about 3 percentage points.

The dataset used in the paper is a de-noising diffusion probabilistic model (DDPM) trained on CIFAR-10, CIFAR-100, STL-10, and Tiny-ImageNet.

Datasets:

Asirra

CIFAR-10

CIFAR-100

GTSRB

HASYv2

MNIST

STL-10

SVHN

The MNIST, MNIST-M, Street View House Numbers (SVHN), Synthetic Digits (SYN DIGITS), CIFAR-10 and STL-10 datasets are used for visual domain adaptation experiments.

This dataset is about book subjects. It has 2 rows and is filtered where the books is Data structures with STL. It features 10 columns including number of authors, number of books, earliest publication date, and latest publication date.