Fueling the Gold Rush: The Greatest Public Datasets for AI

Luke de Oliveira
Startup Grind
Published in
7 min readFeb 11, 2017


It has never been easier to build AI or machine learning-based systems than it is today. The ubiquity of cutting edge open-source tools such as TensorFlow, Torch, and Spark, coupled with the availability of massive amounts of computation power through AWS, Google Cloud, or other cloud providers, means that you can train cutting-edge models from your laptop over an afternoon coffee.

Though not at the forefront of the AI hype train, the unsung hero of the AI revolution is data — lots and lots of labeled and annotated data, curated with the elbow grease of great research groups and companies who recognize that the democratization of data is a necessary step towards accelerating AI.

However, most products involving machine learning or AI rely heavily on proprietary datasets that are often not released, as this provides implicit defensibility.

With that said, it can be hard to piece through what public datasets are useful to look at, which are viable for a proof of concept, and what datasets can be useful as a potential product or feature validation step before you collect your own proprietary data.

It’s important to remember that good performance on data set doesn’t guarantee a machine learning system will perform well in real product scenarios. Most people in AI forget that the hardest part of building a new AI solution or product is not the AI or algorithms — it’s the data collection and labeling. Standard datasets can be used as validation or a good starting point for building a more tailored solution.

This week, a few machine learning experts and I were talking about all this. To make your life easier, we’ve collected an (opinionated) list of some open datasets that you can’t afford not to know about in the AI world.

Legend:📜 Classic — these are some of the more famous, legacy, or storied datasets in AI. It’s hard to find a researcher or engineer who hasn’t heard of them.🛠 Useful — these are datasets that are about as close to real-world that a curated, cleaned dataset can be. Also, these are often general enough to be useful in both the product and R&D world.📚 Academic baseline — these are datasets that are commonly used the in the academic side of Machine Learning and AI as benchmarks or baselines. For better or worse, people use these datasets to validate algorithms.🗿 Old - these datasets, irrespective of utility, have been around for a while.

Computer Vision

  • 📚 📜 🗿 MNIST: most commonly used sanity check. Dataset of 25x25, centered, B&W handwritten digits. It is an easy task — just because something works on MNIST, doesn’t mean it works.
  • 📜 🗿 CIFAR 10 & CIFAR 100: 32x32 color images. Not commonly used anymore, though once again, can be an interesting sanity check.
  • 🛠 📚 📜 ImageNet: the de-facto image dataset for new algorithms. Many image API companies have labels from their REST interfaces that are suspiciously close to the 1000 category WordNet hierarchy from ImageNet.
  • LSUN: Scene understanding with many ancillary tasks (room layout estimation, saliency prediction, etc.) and an associated competition.
  • 📚 PASCAL VOC: Generic image Segmentation / classification — not terribly useful for building real-world image annotation, but great for baselines.
  • 📚 SVHN: House numbers from Google Street View. Think of this as recurrent MNIST in the wild.
  • MS COCO: Generic image understanding / captioning, with an associated competition.
  • 🛠 Visual Genome: Very detailed visual knowledge base with deep captioning of ~100K images.
  • 🛠 📚 📜 🗿 Labeled Faces in the Wild: Cropped facial regions (using Viola-Jones) that have been labeled with a name identifier. A subset of the people present have two images in the dataset — it’s quite common for people to train facial matching systems here.

Natural Language

  • 🛠 📚 Text Classification Datasets (Google Drive Link) from Zhang et al., 2015: An extensive set of eight datasets for text classification. These are the most commonly reported baselines for new text classification baselines. Sample size of 120K to 3.6M, ranging from binary to 14 class problems. Datasets from DBPedia, Amazon, Yelp, Yahoo!, Sogou, and AG.
  • 🛠 📚 WikiText: large language modeling corpus from quality Wikipedia articles, curated by Salesforce MetaMind.
  • 🛠 Question Pairs: first dataset release from Quora containing duplicate / semantic similarity labels.
  • 🛠 📚 SQuAD: The Stanford Question Answering Dataset — broadly useful question answering and reading comprehension dataset, where every answer to a question is posed as a span, or segment of text.
  • CMU Q/A Dataset: Manually-generated factoid question/answer pairs with difficulty ratings from Wikipedia articles.
  • 🛠 Maluuba Datasets: Sophisticated, human-generated datasets for stateful natural language understanding research.
  • 🛠 📚 Billion Words: large, general purpose language modeling dataset. Often used to train distributed word representations such as word2vec or GloVe.
  • 🛠 📚 Common Crawl: Petabyte-scale crawl of the web — most frequently used for learning word embeddings. Available for free from Amazon S3. Can also be useful as a network dataset for it’s crawl of the WWW.
  • 📚 📜 bAbi: synthetic reading comprehension and question answering dataset from Facebook AI Research (FAIR).
  • 📚 The Children’s Book Test (download link): Baseline of (Question + context, Answer) pairs extracted from Children’s books available through Project Gutenberg. Useful for question-answering, reading comprehension, and factoid look-up.
  • 📚 📜 🗿 Stanford Sentiment Treebank: standard sentiment dataset with fine-grained sentiment annotations at every node of each sentence’s parse tree.
  • 📜 🗿 20 Newsgroups: one of the classic datasets for text classification, usually useful as a benchmark for either pure classification or as a validation of any IR / indexing algorithm.
  • 📜 🗿 Reuters: older, purely classification based dataset with text from the newswire. Commonly used in tutorials.
  • 📜 🗿 IMDB: an older, relatively small dataset for binary sentiment classification. Fallen out of favor for benchmarks in the literature in lieu of larger datasets.
  • 📜 🗿 UCI’s Spambase: Older, classic spam email dataset from the famous UCI Machine Learning Repository. Due to details of how the dataset was curated, this can be an interesting baseline for learning personalized spam filtering.


Most speech recognition datasets are proprietary — the data holds a lot of value for the company that curates. Most datasets available in the field are quite old.

  • 📚 🗿 2000 HUB5 English: English-only speech data used most recently in the Deep Speech paper from Baidu.
  • 📚 LibriSpeech: Audio books data set of text and speech. Nearly 500 hours of clean speech of various audio books read by multiple speakers, organized by chapters of the book containing both the text and the speech.
  • 🛠 📚 VoxForge: Clean speech dataset of accented english, useful for instances in which you expect to need robustness to different accents or intonations.
  • 📚 📜 🗿 TIMIT: English-only speech recognition dataset.
  • 🛠 CHIME: Noisy speech recognition challenge dataset. Dataset contains real, simulated and clean voice recordings. Real being actual recordings of 4 speakers in nearly 9000 recordings over 4 noisy locations, simulated is generated by combining multiple environments over speech utterances and clean being non-noisy recordings.
  • TED-LIUM: Audio transcription of TED talks. 1495 TED talks audio recordings along with full text transcriptions of those recordings.

Recommendation and ranking systems

  • 📜 🗿 Netflix Challenge: first major Kaggle style data challenge. Only available unofficially, as privacy issues arose.
  • 🛠 📚 📜 MovieLens: various sizes of movie review data — commonly used for collaborative filtering baselines.
  • Million Song Dataset: large, metadata-rich, open source dataset on Kaggle that can be good for people experimenting with hybrid recommendation systems.
  • 🛠 Last.fm: music recommendation dataset with access to underlying social network and other metadata that can be useful for hybrid systems.

Networks and Graphs

  • 📚 Amazon Co-Purchasing and Amazon Reviews: crawled data from the “users who bought this also bought…” section of Amazon, as well as amazon review data for related products. Good for experimenting with recommendation systems in networks.
  • Friendster Social Network Dataset: Before their pivot as a gaming website, Friendster released anonymized data in the form of friends lists for 103,750,348 users.

Geospatial data

  • 🛠 📜 OpenStreetMap: Vector data for the entire planet under a free license. It includes (an older version of) the US Census Bureau’s TIGER data.
  • 🛠 Landsat8: Satellite shots of the entire Earth surface, updated every several weeks.
  • 🛠 NEXRAD: Doppler radar scans of atmospheric conditions in the US.

❗️People often think solving a problem on one dataset is equivalent to having a well thought out product. Use these datasets as validation or proofs of concept, but don’t forget to test or prototype how the product will function and obtain new, more realistic data to improve its operation. Successful data-driven companies usually derive strength from their ability to collect new, proprietary data that improves their performance in a defensible way.

Please contribute!

If you think we’ve missed a dataset or two (which we definitely have!) or have a conflicting opinion about a dataset discussed here, please let me know with a comment, or you can shoot me an email at lukedeo@ldo.io!

P.S. — This post is part of a open, collaborative effort to build an online reference, the Open Guide to Practical AI, which we’ll release in draft form in a few weeks. See this popular previous guide for an example. If you’d like to get updates on or help with with this effort, drop me a comment or email me at lukedeo@ldo.io.

Special thanks to Joshua Levy, Srinath Sridhar, and Max Grigorev.



Luke de Oliveira
Startup Grind

Deep learning, Infrastructure, and Open Source. Founder @ Vai, Visiting scientist @ Berkeley Labs, Stanford/Yale Alum.