The Hugging Face Hub is the default place to host and download open ML datasets, with over a million of them. It is not the only place to look: dataset cards and licence fields are written by uploaders, finding the right dataset among a million takes knowing what to search for, and plenty of useful data, from government statistics to research archives, is published elsewhere. These fourteen sources cover that ground.
The short list
Disclosure. Dataset Finder publishes this guide and is one of the options in it. We have kept to facts each company states or that were reported in the press, and linked our detailed comparison pages so you can check the differences yourself.
Hugging Face stays the right tool for hosting your own datasets and for pulling a dataset you already know the name of. The alternatives below help when you do not know what exists, need a specific domain, or need terms an open licence does not give you.
| Option | Type | Pricing | Best for |
|---|---|---|---|
| Dataset Finder (ours) | Dataset search + self-serve custom data | Free plan; Startup €9.90/month; Pro €29.90/month. Custom data paid in credits, with a public cost estimator. | Teams that want to check what already exists, then order the gaps self-serve |
| Kaggle Datasets | Dataset hub + notebooks | Free | Exploring a dataset in a notebook with free compute |
| Google Dataset Search | Search engine | Free, no account | Government, scientific and statistical data |
| OpenML | Dataset hub (tabular) | Free | Tabular benchmarking and AutoML research |
| UCI Machine Learning Repository | Repository | Free | Teaching, baselines and classic tabular benchmarks |
| Zenodo | Research repository | Free | Citable research datasets attached to papers |
| Harvard Dataverse | Research repository | Free | Replication data across academic disciplines |
| Registry of Open Data on AWS | Cloud data registry | Free to browse; compute billed by AWS | Very large geospatial, climate and genomics data |
| data.europa.eu | Official open-data portal | Free | European public-sector and statistical data |
| Roboflow Universe | Dataset hub (computer vision) | Free to browse; Roboflow platform from $0 | Quick-start detection and segmentation data |
| TensorFlow Datasets | Dataset library | Free, open source | One-line loading in TensorFlow or JAX |
| Papers with Code | Benchmark index | Free | Finding which datasets a benchmark result used |
| Defined.ai | Data marketplace | Not published; samples and quotes on request | Licensed speech and multilingual data |
| Shaip | Data marketplace + services | Not published | Licensed medical and speech corpora |
What Hugging Face does well, and where it stops
The Hub listed 1,105,943 datasets on 26 October 2026, with free public hosting, the in-browser Dataset Viewer, gated and private datasets and direct integration with the datasets library. In September 2026 NVIDIA agreed to acquire Hugging Face; the deal is pending approvals and expected to close in the first half of 2027, and NVIDIA says Hugging Face will remain an open platform.
The Hub offers keyword and full-text search with filters for modality, task, language, licence and size. Dataset cards, licence fields and descriptions are written by whoever uploaded the dataset, so the information is only as good as the uploader’s. Our Dataset Finder vs Hugging Face page covers the difference between hosting and curation.
Search engines and hubs
These help you find datasets rather than host them. Use them when you know the task but not the dataset.
Dataset Finder (ours)
Dataset search + self-serve custom dataSearch tens of thousands of curated datasets in plain language, then order the data that does not exist yet from Innovatiana’s managed annotators in the same app, paid in credits.
- Each dataset card adds automated analysis on 30+ criteria, including licence, PII, bias and fine-tuning readiness.
- Custom requests are labelled by trained annotators and domain specialists under a dedicated Data Labeling Manager, not an open crowd.
- A test batch usually comes back within about 48 hours; large projects move to an Enterprise quote.
- Launched in August 2026, so the product is new; the annotation work is done by Innovatiana, the data-labelling company behind it.
Kaggle Datasets
Dataset hub + notebooksGoogle’s data-science community: user-uploaded datasets next to free notebooks with CPU, GPU and TPU, competitions and courses.
- Kaggle reported 23M+ accounts in 2025.
- Free notebooks run right next to each dataset, and the kagglehub library downloads and loads datasets from Python.
Google Dataset Search
Search engineA free search engine over dataset metadata published across the web, from government, scientific and ML repositories.
- Indexes metadata published with schema.org or DCAT, not the data itself.
- Came out of beta in January 2020 with almost 25 million datasets indexed.
Research repositories and open-data portals
Academic archives and public-sector portals hold data that rarely reaches ML hubs: replication datasets, tabular benchmarks, government statistics and very large scientific collections. Licences are set by each depositor, so check every record.
OpenML
Dataset hub (tabular)A community-run platform for sharing ML datasets, tasks and experiment results, with Python, R and Java APIs and scikit-learn integration.
- Each dataset carries its own licence; the platform’s terms describe it as free for research, education and non-commercial use.
- Computes data-quality meta-features for every dataset automatically.
UCI Machine Learning Repository
RepositoryOne of the oldest ML dataset archives, run at UC Irvine since 1987, mostly small and medium tabular datasets.
- UCI says it maintains 689 datasets.
- Dataset pages state a licence, and the ucimlrepo package loads them in Python.
Zenodo
Research repositoryA free open research archive built by CERN and OpenAIRE, where researchers deposit datasets, code and papers with citable DOIs.
- Records hold up to 50GB, with higher quotas granted case by case.
- Uploaders set the licence for each record, so terms vary.
Harvard Dataverse
Research repositoryA free research data repository open to researchers in any discipline, built on the open-source Dataverse software.
- CC0 is the recommended default for public datasets, but owners can set other terms or restrict access.
- Researchers can archive up to 1TB at no cost.
Registry of Open Data on AWS
Cloud data registryA catalogue of public datasets stored on AWS, provided by organisations such as NASA, NOAA and NIH.
- Datasets are maintained by third parties under a variety of licences, so each listing must be checked.
- Storage is sponsored for approved datasets; you pay for your own compute, and some buckets are Requester Pays.
data.europa.eu
Official open-data portalThe official portal for European data, run by the Publications Office of the EU, aggregating public-sector catalogues.
- Lists over 1.8 million datasets from 213 catalogues across 36 countries.
- Licences are set by each publisher; a Licensing Assistant compares their terms.
Domain hubs and loaders
For computer vision, Roboflow says its Universe hub holds over 1 million datasets. TensorFlow Datasets is a loader rather than a host. Papers with Code, once the standard way to find the datasets behind a benchmark, was shut down in July 2025; a revival led by Hugging Face staff now runs at paperswithcode.co.
Roboflow Universe
Dataset hub (computer vision)Roboflow’s public hub of user-contributed computer-vision datasets and fine-tuned models.
- Roboflow says Universe holds over 1 million datasets and over 250,000 fine-tuned models.
- Licences are set per dataset by contributors, so quality and terms vary.
TensorFlow Datasets
Dataset libraryGoogle’s open-source Python library that downloads and prepares public datasets as ready-made pipelines for TensorFlow and JAX.
- The library does not host the data; checking each dataset’s licence is your responsibility.
- Still maintained: version 4.9.10 was released in May 2026.
Papers with Code
Benchmark indexThe index that linked ML papers to code, datasets and leaderboards. The original Meta-run site was shut down in July 2025.
- Users report that paperswithcode.com now redirects to Hugging Face’s Trending Papers, which Hugging Face staff called only a partial alternative.
- The legacy data dumps remain available under CC BY-SA 4.0, and a revival led by Hugging Face staff runs at paperswithcode.co.
When you need licensed data
Open licences do not cover every use. If you need a commercial licence for speech, medical or multilingual corpora, data marketplaces sell ready-made datasets with the rights spelled out. If the data does not exist at all, the alternative is to commission it: Dataset Finder’s custom requests are labelled by Innovatiana’s managed annotators, and the cost estimator gives a first figure.
Defined.ai
Data marketplaceLicensable off-the-shelf datasets, especially speech and multilingual, plus custom collection and annotation.
- Lists ISO 42001 alongside ISO 27001 and ISO 27701.
- Reported 65% revenue growth for 2025.
Shaip
Data marketplace + servicesData collection, annotation and licensing, strong in medical and speech data; part of Ubiquity since 2026.
- Shaip’s homepage cites 30M patient notes and 250k hours of medical audio.
- Ubiquity Global Services agreed to acquire it in February 2026.
Before you download from any source
- Read the licence on the source, not the aggregator. A hub’s licence field is metadata someone typed; the terms that bind you are the publisher’s. See dataset licences explained.
- Check provenance. Derived datasets inherit the terms of what they were built from. Our provenance guide lists what to trace.
- Look at the data, not the card. Sample a few hundred rows for label noise, duplicates and leakage before you train.
- Keep a datasheet. Where the data came from and under which terms is easier to record on download than to reconstruct later; the EU AI Act datasheet generator structures it.
Sources
Facts were checked on 26 October 2026 against each company’s own site, press releases and filings, and reputable press for corporate events; figures a company reports about itself are marked as its own. Main sources: Dataset Finder, OpenML, UCI Machine Learning Repository, Zenodo, Harvard Dataverse, Registry of Open Data on AWS, data.europa.eu, Roboflow Universe, TensorFlow Datasets, Papers with Code. The comparison pages list every source for the vendors they cover.
Questions people ask
What is the best alternative to Hugging Face datasets?
It depends on the job. Kaggle for exploring data with free notebooks, Google Dataset Search for government and scientific data, OpenML or UCI for tabular benchmarks, Roboflow Universe for computer vision, and Dataset Finder for searching a curated catalogue by what you are building.
Is Papers with Code still available?
The original Meta-run site was shut down in July 2025, and users report that paperswithcode.com now redirects to Hugging Face’s Trending Papers. The legacy data dumps remain available under CC BY-SA 4.0, and a revival led by Hugging Face staff runs at paperswithcode.co.
Are datasets on these sites free to use commercially?
Not automatically. Each dataset carries its own licence set by whoever published it, and some repositories describe the platform itself as for research and non-commercial use. Check the licence on the original source before training a commercial model.
Is Hugging Face being acquired?
NVIDIA signed an agreement to acquire Hugging Face in September 2026. It is subject to regulatory approvals and expected to close in the first half of 2027; NVIDIA says Hugging Face will remain an open platform.