19 Dataset Repositories & Sources
-
01UCI Machine Learning Repository
The University of California Irvine Machine Learning Repository is one of the oldest and most widely cited collections in the field — hosting over 600 curated datasets spanning classification, regression, clustering, and anomaly detection tasks. Each dataset entry includes standardized metadata: attribute types, class distributions, missing value information, and citation counts, making it the go-to resource for academic benchmarking and reproducible ML research.
Best for: Researchers and students evaluating supervised and unsupervised learning algorithms on standardized, citable benchmark datasets — the gold standard for ML academic reproducibility. -
02GitHub — 650+ Public API Datasets
This curated GitHub repository catalogs over 650 public APIs across categories including weather, finance, transportation, social media, geospatial, sports, and science — each providing programmatic access to real-world, continuously updated data suitable for ML feature engineering and pipeline building. It serves as a comprehensive starting point for developers who need live data sources rather than static training files, enabling models trained on real-world signal rather than historical snapshots.
Best for: Engineers building data pipelines and feature stores who need live, domain-specific real-world data from public APIs — particularly for time series, financial, weather, and geospatial ML applications. -
03PMLB — Penn Machine Learning Benchmarks
The Penn Machine Learning Benchmarks (PMLB), curated by the Epistasis Lab, provides a standardized collection of benchmark datasets specifically designed for rigorous, reproducible empirical evaluation of supervised learning algorithms. The collection covers binary classification, multi-class classification, and regression problems — with a Python API enabling programmatic dataset loading and direct integration with scikit-learn workflows for automated algorithm comparison.
Best for: ML researchers and AutoML practitioners running systematic algorithm comparisons across hundreds of standardized datasets — PMLB's uniform API and broad coverage make it the benchmark suite of choice for empirical ML studies. -
04Papers With Code — 3,250 ML Datasets
Papers With Code aggregates over 3,250 machine learning datasets alongside their associated research papers, evaluation leaderboards, and State-of-the-Art (SoTA) benchmark results — creating a unified view of what data is used and how models perform on it. Its rich metadata schema enables discovery by task type, modality (text, image, audio, video, graph), dataset size, and licensing — directly linking each dataset to the published methods evaluated on it.
Best for: Researchers tracking State-of-the-Art results and finding datasets that align with specific ML tasks — Papers With Code uniquely links datasets to benchmark leaderboards and the papers that introduced or used them. -
05OpenML — Open Machine Learning Platform
OpenML is an open, collaborative machine learning platform hosting thousands of datasets, standardized ML tasks, and experimental results — all accessible through a REST API and Python/R/Java clients. It integrates natively with scikit-learn, Weka, and mlr3, enabling automated experiment tracking, reproducibility of published results, and community sharing of ML workflows. OpenML's task framework standardizes evaluation setups (train/test splits, evaluation metrics) across contributed experiments.
Best for: ML researchers and AutoML practitioners who want programmatic access to thousands of tasks and experiments — OpenML's API integration with scikit-learn makes it ideal for large-scale automated benchmarking and meta-learning research. -
06VisualData Discovery
VisualData Discovery is a specialized search engine and curated directory for computer vision datasets — spanning object detection, image segmentation, facial recognition, medical imaging, satellite imagery, autonomous driving, and action recognition. Its taxonomy and filter system enable CV researchers to quickly identify the right dataset by modality, annotation type, image count, and licensing — covering both open-access academic datasets and commercially usable corpora.
Best for: Computer vision researchers and ML engineers needing to discover CV-specific datasets by task type and annotation format — VisualData's CV-first taxonomy surfaces specialized datasets that general search tools frequently miss. -
07Roboflow Public CV Datasets
Roboflow's public dataset hub provides over 200,000 computer vision datasets spanning real-world domains — from industrial defect inspection to wildlife tracking, retail shelf analysis, and medical imaging — with pre-built export formats for YOLO, COCO, TensorFlow, and Pascal VOC. Beyond raw data, Roboflow offers built-in annotation tools, data augmentation pipelines, and dataset versioning — making it a complete end-to-end platform for CV model development from data collection to training.
Best for: Computer vision practitioners needing labeled, pre-formatted training data for YOLO/COCO-style object detection — Roboflow's annotation tools, augmentation, and version control make it the most complete CV dataset workflow platform available. -
08Kaggle Datasets
Kaggle hosts one of the world's largest community-driven collections of machine learning datasets — covering NLP, computer vision, tabular data, time series, and domain-specific datasets from healthcare, finance, e-commerce, and sports. Its integrated search, filter, and Python download API, combined with built-in Jupyter notebooks and competition leaderboards, make Kaggle the de facto hub for both learning ML and sourcing real-world training data for production applications.
Best for: Data scientists at every level — from beginners learning ML on competition datasets to professionals sourcing labeled training data for production models across virtually every industry vertical. -
09Jupyter Datasets Reference Guide
The Jupyter Tutorial documentation dataset page aggregates commonly available datasets and data search engines organized by domain — serving as a concise reference guide for data scientists working in Jupyter notebooks. It compiles quick-access links to classic ML benchmark datasets (Iris, Boston Housing, MNIST) alongside pointers to broader dataset discovery platforms, making it an ideal starting point for newcomers setting up their first data science environment.
Best for: Data science students and Jupyter notebook users looking for a quick, organized reference to classic ML datasets and dataset search engines — ideal as a starting point before exploring larger repositories. -
10IBM Data Asset eXchange (DAX)
IBM's Data Asset eXchange (DAX) provides a curated collection of enterprise-grade open datasets for data science — including audio, image, text, and structured datasets — all pre-vetted for quality, provenance, and clear licensing terms. Each dataset on DAX ships with associated Jupyter notebooks and example model training code, dramatically lowering the barrier to ML experimentation. DAX prioritizes datasets that represent real-world enterprise challenges across healthcare, finance, climate, and business domains.
Best for: Enterprise data scientists and IBM Cloud/Watson Studio users seeking high-quality, pre-licensed datasets for production ML use cases — DAX's notebook+dataset bundles are ideal for rapid prototyping with clear IP provenance. -
11Google Dataset Search
Google Dataset Search is a specialized search engine that indexes dataset repositories across the web by crawling Schema.org Dataset markup — surfacing datasets from government open data portals, academic repositories, cloud provider catalogs, and research institutions. It enables researchers to find datasets by topic, time period, update frequency, and usage licensing with Google's full-text search relevance — functioning as a "Google Scholar for datasets."
Best for: Researchers needing to find datasets across the entire open web — Google Dataset Search discovers domain-specific and governmental datasets that are not indexed on Kaggle or UCI, making it essential for niche research domains. -
12GitHub — Awesome Public Datasets
Awesome Public Datasets is a community-maintained GitHub repository curating high-quality open datasets organized by domain — covering agriculture, biology, climate, economics, finance, government, healthcare, machine learning, music, natural language, and more. With tens of thousands of GitHub stars, it is a standard reference for ML practitioners and data engineers seeking domain-specific real-world training data outside mainstream ML-focused repositories.
Best for: ML practitioners and domain experts needing public datasets in specific industries — Awesome Public Datasets' taxonomy covers niche domains (agriculture, climate, social science) underrepresented on Kaggle or UCI. -
13DBpedia — Open Knowledge Graph
DBpedia is a community-driven initiative that extracts structured, machine-readable facts from Wikipedia and Wikimedia projects — producing a massive open knowledge graph covering hundreds of millions of facts about people, places, organizations, creative works, and scientific concepts. It is a foundational dataset for knowledge-graph ML tasks, entity linking, semantic search, relation extraction, and as enrichment data for training knowledge-grounded NLP and question-answering models.
Best for: NLP researchers and knowledge graph engineers needing a large, well-structured open knowledge base — DBpedia's Wikipedia derivation gives it unmatched coverage and community maintenance for entity-level ML tasks. -
14WordNet — Princeton Lexical Database
WordNet is a large lexical database of English developed by Princeton University — grouping nouns, verbs, adjectives, and adverbs into sets of cognitive synonyms (synsets) each expressing a distinct concept, linked by semantic relations including hypernymy (is-a), holonymy (part-of), meronymy, and antonymy. It is a cornerstone resource for NLP tasks including word sense disambiguation, semantic similarity computation, taxonomy construction, and as the hierarchical backbone of the ImageNet image database.
Best for: NLP researchers building semantic similarity models, word sense disambiguation systems, and ontology-grounded language models — WordNet's synset hierarchy is foundational to both classical NLP and modern transformer fine-tuning. -
15FactForge — Linked Open Data Hub
FactForge.net (by Ontotext) is a hub of Linked Open Data combining structured knowledge from DBpedia, GeoNames, Freebase, and other authoritative sources with real-time news articles about people, organizations, and locations. It provides SPARQL-queryable access to billions of cross-linked facts for semantic ML, knowledge graph completion, entity resolution, and news-aware entity analytics — making it valuable for researchers working at the intersection of structured knowledge and unstructured text.
Best for: Semantic AI and knowledge graph researchers who need cross-linked LOD resources queryable via SPARQL — FactForge's integration of multiple authoritative sources makes it ideal for entity resolution and knowledge graph completion tasks. -
16World Facts Database
World Facts provides a structured, machine-readable database of authoritative reference data on nations, languages, currencies, time zones, capital cities, populations, geographic coordinates, and calling codes. It serves as a reliable enrichment dataset for geospatial NLP tasks, internationalization feature engineering, country-level ML aggregations, and as a lookup table for models requiring geographic context about entities referenced in text or structured data.
Best for: Data engineers enriching ML feature sets with geospatial, demographic, and national reference data — particularly for global-scale NLP entity normalization, currency/language detection, and geographic feature engineering. -
17GLEIF — Global Legal Entity Identifier Foundation
The Global Legal Entity Identifier Foundation (GLEIF) provides the world's authoritative open database of legal entities — assigning unique LEI codes to businesses and organizations across 200+ jurisdictions with standardized, continuously maintained reference data. The GLEIF dataset is the gold standard for financial ML tasks including entity resolution, KYC (Know Your Customer) model training, fraud detection, supply chain graph analytics, and counterparty risk modeling — with free bulk download and API access.
Best for: Financial ML engineers and RegTech practitioners building entity resolution, AML/fraud detection, or KYC models — GLEIF's LEI data provides uniquely authoritative, globally standardized legal entity reference data unavailable elsewhere. -
18ConceptNet — Semantic Network
ConceptNet is a freely available multilingual commonsense semantic network — linking words and phrases across 80+ languages with labeled relationships including "IsA", "UsedFor", "PartOf", "AtLocation", "CapableOf", and "Causes". Built from expert-curated knowledge, Open Mind Common Sense crowdsourcing, and Wiktionary extraction, ConceptNet is widely used for commonsense reasoning AI, NLP transfer learning enrichment, dialogue systems, and question-answering models that require world knowledge beyond what appears in training corpora.
Best for: NLP and AI researchers building commonsense reasoning systems, dialogue agents, and knowledge-augmented language models — ConceptNet's labeled relation types and multilingual coverage are uniquely suited to grounding language models in real-world knowledge. -
19ImageNet
ImageNet is the benchmark image database that defined modern deep learning for computer vision — organizing over 14 million labeled images into 20,000+ synsets according to the WordNet noun hierarchy, each node depicted by hundreds to thousands of images. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) drove breakthrough advances in deep CNN architectures — AlexNet (2012), VGG, GoogLeNet/Inception, ResNet — establishing ImageNet as the foundational pre-training corpus for transfer learning in virtually all computer vision applications.
Best for: Deep learning researchers and CV engineers — ImageNet pre-training weights (via PyTorch torchvision, TensorFlow Hub) power transfer learning for image classification, object detection, and segmentation in nearly every production CV application today.