The stack has largely settled: pandas or Polars for data, scikit-learn for classical models, PyTorch for deep learning. The real skill here is not picking a model but cleaning the data, which is where most of the work goes.
83 links
A community archive distributing large research datasets over BitTorrent. Against institutional download links the difference is durability: when the university server goes away, the data survives in the swarm.
The standard introduction to statistical learning, free as a PDF with separate R and Python labs. Against "Elements of Statistical Learning" by the same authors it trades mathematics for intuition — this is the one you read first.
A competition on abstract reasoning whose tasks are designed to be easy for humans and hard for models. It rewards a new approach rather than more scale.
Developer resources for ArcGIS, the mapping and location-analysis platform. Unlike open-source GIS lists, it centres on the commercial Esri ecosystem, and Esri maintains it.
A list of ready machine-learning models for Apple’s Core ML framework. Unlike general model hubs, every model is already in Core ML format for on-device use on iOS/macOS; part of the list is dated.
A list of tools and resources for CSV, the tabular-data file format. Unlike strictly schema’d formats (Parquet, JSON Schema) CSV is loose, and the list gathers the command-line tools and libraries that cope with its quirks.
A list of courses, papers, frameworks and datasets for deep learning. Unlike general ML lists it focuses on neural networks and modern architectures.
A list compiling the most-cited deep-learning papers of 2012–2016; it is no longer maintained. Unlike live trackers such as Papers with Code, it offers a fixed core chosen by citation count — a historical entry point to the field’s foundations.
A list for geographic information systems: cartographic tools, geoanalysis libraries, data sources and communities. Unlike Esri-centred lists it is open-source and platform-neutral, with desktop GIS, web mapping libraries and open data in one place.
A list for H2O, the open-source distributed machine-learning platform. Unlike scikit-learn, it runs distributed across a cluster and has AutoML built in.
A list for JAX, which offers automatic differentiation and XLA compilation. Unlike PyTorch it is functional: you apply transformations such as grad, jit and vmap to pure functions.
A list of tools and resources for JSON, the text-based data-interchange format. Unlike a single parser, it brings together schema validators, query tools such as jq and libraries per language.
A resource list for Jupyter notebooks, which hold code, text and charts together. Unlike the Jupyter docs, it gathers the ecosystem around it — extensions, kernels, tools that turn a notebook into an app or slides.
A list for generating text from data or dialogue (natural-language generation). Unlike general NLP lists it is about producing text rather than understanding it; much of it predates large language models.
A list of tools and resources for OpenStreetMap, the open map-data project. Unlike the Google Maps API the data is openly licensed (ODbL) and fully downloadable; the list gathers editors, processors and renderers.
A list of public datasets organised by topic. Unlike random data search it points to clean, usable sources grouped by domain.
A list organising Python data-science libraries by workflow (data, models, visualisation). Unlike a PyPI search it shows curated tools for each stage.
A list for asking and answering questions in natural language (QA). Unlike general NLP lists, it focuses only on question answering — datasets, models, papers.
Resources for machine-learning engineering from experiment to production. Unlike model lists it looks at the “making ML a product” side — testing, deployment, monitoring.
A list of resources and models for TensorFlow, the machine-intelligence library. Unlike the official docs, it gathers community tutorials, models and projects; part of it dates from the TensorFlow 1 era.
A registry of large public datasets hosted on AWS. Unlike portals such as Data.gov the data sits directly in S3, so you can process it inside AWS without downloading.
A reference giving geography, population, economy, government and military indicators for every country on uniform pages; public domain. Unlike Wikipedia country articles every country has the same fields in the same order, which makes comparison easy. The publisher is an intelligence agency — keep that in mind for interpretive fields.
A Google-backed project that merges statistics from the Census, WHO, the UN and dozens of other sources into one knowledge graph, queryable in natural language and through an API. Unlike single-agency portals it sets series from different sources side by side in one query. Google’s old Public Data Explorer address redirects here.
Microsoft’s 20-lesson curriculum, each unit with pre/post quizzes and a project. Unlike most technical courses, it gives data ethics and visualisation sections of their own.
The European Union’s official open-data portal. Unlike national portals, it merges member states’ datasets into one search and also includes data from the EU institutions themselves.
A directory of machine learning datasets by domain, with size, licence and paper links. Against Kaggle these are academic and industrial sets rather than competition data — for when you need something real.
Andrew Ng’s education platform, with specialisations in deep learning and generative AI. Unlike the long specialisations, its free one-hour short courses are built with industry partners.
Public datasets that live inside BigQuery: GitHub, Wikipedia, weather, the census. Unlike the AWS Open Data Registry the data comes as tables you query with SQL rather than files; within the monthly free query quota you can work without downloading.
Searches open datasets scattered across the web via schema.org markup. Unlike a single repository’s search it is tied to none, so it reaches institutional archives too.
A tool showing how search interest changes over time and place. Unlike keyword planners it gives a relative 0–100 index rather than absolute search counts; useful for gauging public interest, seasonality and rising topics.
Crisis and humanitarian datasets curated by OCHA. Unlike general open-data portals, it focuses on crisis areas, with data coming from organisations in the field.
Bite-sized micro-courses covering Python, Pandas, SQL, and machine learning through in-browser notebook exercises; focuses on hands-on coding tasks rather than long video lectures.
Active and past competitions. Unlike working on your own project, you are scored against a real metric on a hidden test set — which breaks the “looks fine to me” illusion.
An open-source tool, born at Uber, for visualising large location datasets (millions of points, trips, grids) layer by layer in the browser. Unlike QGIS it aims at exploration and presentation rather than analysis: drop in a CSV or GeoJSON and you have a map with a time slider in minutes.
A book repo that has you write a GPT from scratch in PyTorch — tokeniser, attention, training loop, fine-tuning. You build the mechanism instead of calling a library.
Application-focused ML tutorials, each ending in working code — answering “how do I do this” rather than the theory.
Interactive WebGL map clustering subreddits by shared commenters (2020-21 comments). Unlike subreddit lists, it shows how communities relate spatially.
The foundation of Python plotting — seaborn and pandas’ plotting are built on it. Unlike interactive libraries such as Plotly it is static but gives fine control, at the cost of considerable code even for a simple chart.
Google’s fast introduction to machine learning, moving through short video, reading and runnable exercises. Compared with long specialisations it is done in a few days, and free.
An open machine learning course. Unlike lecture-only material, it pairs theory with Kaggle competition assignments.
Amazon’s set explaining machine learning concepts through interactive visuals. Unlike a textbook, you drag a parameter and watch the model change for yourself.
NASA’s portal to its satellite and Earth-observation data, from rainfall and ice cover to vegetation and the atmosphere. Unlike NOAA’s ground-station records it rests on measurements taken from space; downloads need a free Earthdata account.
A free, open-source weather API that needs no key for non-commercial use. Unlike commercial weather APIs it asks for no signup, and its data is open.
One of the largest open databases of companies, merging national company registers into one search. Unlike individual national registers it presents many jurisdictions in the same form; officer and affiliate links make it useful for investigation, while heavy API use is paid.
A platform institutions and cities use to run data portals, with datasets searchable across portals. Rather than visiting portals one by one, you can search the data of every portal built on it together.
X-lab's analysis platform publishing open-source health metrics derived from GitHub activity (OpenRank, activity, contributors). Unlike the raw GitHub API, the metrics come ready-made as open time-series data.
A web front end for querying OpenStreetMap data through the Overpass API; it shows the result on a map, downloads it as GeoJSON and has a wizard for those who do not write queries. Unlike downloading a full OSM extract it fetches only the features you ask for, such as every pharmacy in a city.
A site pairing machine-learning papers with their implementation code and benchmark tables. Unlike a paper alone it puts the result’s code and the leaderboard beside it.
An open-source desktop GIS for Windows, macOS and Linux; it reads and analyses shapefile, GeoPackage, PostGIS and raster layers and extends through plugins. Unlike ArcGIS there is no licence fee; in exchange support comes from the community and commercial partners.
A community for finding and requesting datasets. Unlike catalogue sites you can post a request when the data you want is missing and see what others have found; there is no quality control, so check the licence and source yourself.
Free lecture notes on the scientific Python stack — NumPy, SciPy, Matplotlib. Unlike single-library docs they show how the pieces fit together.
The standard Python library for classical machine learning — classification, regression, clustering and preprocessing behind one fit/predict interface. Unlike deep-learning frameworks such as PyTorch, it is a ready, consistent toolbox for tabular data and classical methods.
A framework turning any video game into a reinforcement learning environment — working against real games rather than simulations is what makes it unusual. The repository is archived (last update November 2022); the games and library versions it supports may be out of date.
Google’s deep learning framework. Compared with PyTorch it lost research ground, yet remains strong for mobile and embedded deployment (TF Lite) and production serving (TF Serving).
Shows a language model’s whole training pipeline in working code, from data download to distributed training. Unlike theoretical write-ups, it also covers the infrastructure side they usually skip.
One of the oldest archives for machine-learning research, home to classic small sets such as Iris and Adult. Unlike Kaggle it is an academic archive rather than a contest site: each set comes with its donor, field descriptions and citation, and papers use them as benchmarks.
Uppsala University’s data programme on armed conflict and organised violence, keeping event-level records with coordinates. Unlike news archives each event is verified against coding rules; it is a standard reference in conflict research.
The World Health Organization’s country-level health indicators: causes of death, vaccination, health spending, risk factors. Unlike national health agencies’ figures, it puts countries on the same definitions and methods so they can be compared; the data is also available through an API.
End-to-end ML notebooks walking a working pipeline from data cleaning to model evaluation, progressing through real code rather than concept prose.