Datasets & Tools Reference

This page is the toolbox chapter of the Machine Learning Handbook: where to find data, what tools to use for annotation, what libraries for modeling, how to version and engineer data, and "which dataset to pick for your next exercise".

Data is the starting point of machine learning — and also its ceiling: feature quality determines the model's upper bound, and features come from the data itself. The Glossary teaches you to read the concepts; this page tells you "what tools you should have on hand and where to get them". Recommended to read alongside Data and Data Engineering and Feature Engineering.

How to use

This section is a quick-reference catalog. When doing exercises, first flip to the "Dataset Quick-Reference" to pick data, then download from the "Acquisition Channels", and model with the libraries in the "Common Libraries Toolbox"; as data accumulates, use the "Data Engineering Tools" to manage versioning, quality, and deployment.

Dataset Quick-Reference

All datasets below are real; scale and purpose are as described by their official sources. The "How to Access" column gives the official source; those marked with ⭐ are also built into scikit-learn / Hugging Face, loadable with a single import or one line of load_dataset.

Classification Tasks

Dataset	Scale & Structure	Typical Use	How to Access
Iris ⭐	150 samples, 4 features, 3 classes (setosa/versicolor/virginica)	Introductory binary/multi-class, KNN, logistic regression teaching	scikit-learn load_iris(); UCI
Titanic	891 training records, ~12 features (with missing values)	Introductory binary classification + full feature engineering exercise	Kaggle competition
Adult (Census Income) ⭐	48,842 samples, 14 features	Binary income classification, class imbalance, fairness research	UCI; OpenML ID 1590
Wisconsin Breast Cancer ⭐	569 samples, 30 features	Introductory binary classification, model comparison	scikit-learn load_breast_cancer(); UCI
Wine Quality	Red 1,599 + White 4,898 samples, 11 chemical features	Dual exercise: multi-class / regression	UCI
Credit Card Fraud	284,807 transactions, only 492 fraudulent (0.17%)	Highly imbalanced classification, cost-sensitive learning	Kaggle dataset
Heart Disease	303 samples, 14 features	Small-sample binary classification, cross-validation practice	UCI

Regression Tasks

Dataset	Scale & Structure	Typical Use	How to Access
California Housing ⭐	20,640 samples, 8 features	Introductory regression, linear model teaching	scikit-learn fetch_california_housing()
House Prices (Kaggle)	1,460 training samples, 79 features	Real-world house price regression, feature engineering competition	Kaggle competition
Boston Housing	506 samples, 13 features	Classic regression teaching (has historical controversy, removed from scikit-learn 1.2, use the OpenML version)	OpenML ID 531
Medical Costs	1,338 samples, 7 features	Insurance cost prediction, categorical feature encoding	Kaggle dataset

The Lesson of Boston Housing

Boston Housing was removed from scikit-learn 1.2 because it contains a feature highly correlated with housing price ("proportion of Black population") and is outdated; the academic community generally recommends replacing it with California Housing. It reminds us: the ethics and timeliness of datasets are themselves part of data engineering.

Image Tasks

Dataset	Scale & Structure	Typical Use	How to Access
MNIST (Handwritten Digits) ⭐	70,000 28×28 grayscale images, 10 classes	Vision intro, the "Hello World"	Original source; Kaggle, Hugging Face
Fashion-MNIST ⭐	70,000 28×28 grayscale images, 10 clothing classes	A harder alternative to MNIST	GitHub; Hugging Face
CIFAR-10 / CIFAR-100	60,000 32×32 color images each (10 classes / 100 classes)	CNN intro, data augmentation exercise	Official page; Hugging Face
SVHN (Street View House Numbers)	Over 600,000 32×32 color digit images	Real-world OCR, transfer learning	Stanford UFLDL
Oxford-IIIT Pets	7,349 images, 37 breeds	Small-data image classification / segmentation exercise	VGG official page
ImageNet (ILSVRC subset)	~1.28M training images, 1,000 classes	Large-scale pretraining and model benchmarks	Official site
COCO (Common Objects in Context)	~330K annotated images, 80 object classes, 1.5M instances	Detection, segmentation, keypoint, image captioning	Official site

Text & NLP

Dataset	Scale & Structure	Typical Use	How to Access
IMDB Reviews ⭐	50,000 entries (25k train / 25k test), binary sentiment classification	Sentiment analysis intro, word embeddings	Stanford official; Hugging Face
SST-2 (GLUE sub-task) ⭐	67,349 training sentences	Sentence-level binary sentiment classification, BERT evaluation	GLUE benchmark; Hugging Face
AG News ⭐	~120K training entries, 4 topic classes	Text classification, baseline comparison	Hugging Face
SQuAD (Extractive QA)	Version 2.0: ~150K questions, 50K Wikipedia articles	Extractive reading comprehension, BERT fine-tuning	SQuAD official; Hugging Face
GLUE / SuperGLUE benchmark	Multi-task suite (including linguistic acceptability, entailment, coreference resolution, etc.)	Universal evaluation of pre-trained models	GLUE / SuperGLUE
THUCNews (Chinese News)	740K Chinese news articles, 14 topic classes	Chinese text classification	GitHub mirror; Hugging Face

Recommendation Systems

Dataset	Scale & Structure	Typical Use	How to Access
MovieLens	Three scales: 100K / 1M / 25M	Collaborative filtering, matrix factorization teaching standard	GroupLens official
Netflix Prize	100M ratings (original data no longer available, use mirrors)	Historical benchmark for recommendation competitions	Kaggle mirror
Amazon Reviews	Over 230M reviews, split across multiple domains	Review-based recommendation, implicit feedback	UCSD official
Last.fm Listening Records	User-track implicit interactions	Implicit feedback recommendation, graph-based methods	GroupLens page

Time Series

Dataset	Scale & Structure	Typical Use	How to Access
Air Passengers	144 monthly observations, 1949–1960	Time series intro, ARIMA / naive baseline	Built into R datasets package / statsmodels
ETTh1/ETTh2 (Transformer Temperature)	2 years, every 15 min / hourly, 7 variables	Long-sequence prediction, Transformer time-series benchmark	ETDataset GitHub
Jena Climate	Every 10 min starting from 2003, 14 meteorological variables	Multi-variable time series, LSTM/Transformer example	Keras example page; Kaggle mirror
M5 Forecasting	Historical sales for 30,490 Walmart products	Multi-level sales forecasting competition	Kaggle competition
Stock Prices (Yahoo Finance)	Historical OHLCV data for any ticker	Financial time series (note: for practice only, not trading)	Pull with yfinance

The "Last Mile" of Time Series Data

Time series data inherently carries temporal order — never shuffle it randomly. Splits must use TimeSeriesSplit. When practicing with Air Passengers, start with a naive baseline like "previous value's prediction" — any model must beat that baseline before its value can be discussed.

Dataset Acquisition Channels

Channel	URL	Characteristics	Best For
Kaggle	kaggle.com	Competitions + datasets + Notebook, all in one; active community, rich solution write-ups	Competition practice, real business data, training a "competition mindset"
Hugging Face Datasets	huggingface.co/datasets	One-line load_dataset() loading; seamless integration with Transformers; hosts community datasets	NLP/CV model fine-tuning, rapid prototyping
UCI ML Repository	archive.ics.uci.edu	The oldest teaching data warehouse, 600+ classic datasets	Teaching classics (Iris, Adult, Wine, etc.)
Papers with Code	paperswithcode.com	Binds papers + code + benchmark scores; dataset pages include SOTA leaderboards	Reproducing papers, checking current state-of-the-art for a task
OpenML	openml.org	Datasets come with stable IDs and metadata, directly integrable with scikit-learn	Reproducible experiments, research benchmarks
Google Dataset Search	datasetsearch.research.google.com	Web-wide dataset search engine	Finding niche / domain-specific data
Awesome Public Datasets	GitHub	Open-source data lists organized by domain	Browsing by industry (finance, healthcare, geography…) for inspiration

The Right Way to Get Data from Kaggle

Beyond official competition data, Kaggle has a large number of user-uploaded datasets. Assess quality by three criteria: is the intended use clearly stated, has it been recently updated, and do notebooks contain validation scripts. Popular competitions (Titanic, House Prices) come with official evaluation metrics and leaderboards — the best "practice problems with known answers".

Annotation & Synthetic Data Tools

Supervised learning relies on labels, and manual annotation is expensive and slow. Tools fall into three categories: manual annotation platforms, weak supervision / automatic error correction, and synthetic data generation.

Manual Annotation

Tool	URL	One-Liner
Label Studio	labelstud.io	Open-source annotation platform covering images/text/audio/video/time series, supports multi-person collaboration and ML-assisted annotation, the top choice for personal practice
Labelbox	labelbox.com	Enterprise-grade annotation and data management platform, mature for CV scenarios
Prodigy	prodi.gy	Programmable annotation tool by Explosion, deeply integrated with spaCy/Transformers, great for NLP active learning
Scale AI	scale.com	Commercial crowdsourced annotation service, commonly seen in large-scale data like autonomous driving
AWS SageMaker Ground Truth	official page	Cloud-hosted annotation, supports built-in human review + automatic annotation pipelines

Weak Supervision & Label Correction

Tool	URL	One-Liner
Snorkel	snorkel.org	Programmatic weak supervision framework: write "labeling rules" as functions, automatically fuse to produce noisy training labels
Cleanlab	cleanlab.ai	Uses confident learning to automatically discover and correct label errors in datasets
Argilla	argilla.io	Open-source data annotation and feedback tool for the LLM era, enables "review annotation" of model predictions

Synthetic Data

Tool	URL	One-Liner
SDV (Synthetic Data Vault)	docs.sdv.dev	Generates new data from real data using generative models, preserving distribution and privacy
Faker	faker.readthedocs.io	Generates "realistic-looking" fake data like names/addresses/emails, extremely convenient for tabular/test data
Mimesis	mimesis.name	A high-performance alternative to Faker, supporting multiple languages
scikit-learn Synthesizers	official docs	make_classification / make_regression / make_blobs — generate data with specified shape and difficulty in one line, a teaching gem

When Not to Rely on Synthetic Data

Synthetic data can never replace the "irregularities" in real data — missing values, dirty values, long tails, hidden biases are precisely the most valuable parts of real-world data. Synthetic data is suitable for: privacy-constrained scenarios, supplementing long-tail classes, and testing code correctness. For learning feature engineering, always use real data (see Feature Engineering).

Common Libraries Toolbox

Ordered by usage frequency, with one sentence of "when to use it" per category.

Data Processing & Visualization

• NumPy — the foundation of numerical computing in Python: multi-dimensional arrays and vectorized operations, the bedrock of every library.
• pandas — the de facto standard for tabular data, DataFrame for cleaning, pivoting, grouping, and joining.
• Polars — a faster DataFrame implemented in Rust, several times faster than pandas at 100M+ rows, with a more modern API.
• Dask — pandas' big-data twin: use it for distributed/lazy computation when data exceeds memory.
• matplotlib — the most basic plotting library, can plot anything but verbose API.
• seaborn — a statistical plotting library built on matplotlib, one-liners for distribution/correlation plots.

Traditional Machine Learning

• scikit-learn — the standard library for traditional ML: models, preprocessing, cross-validation, metrics, and Pipeline in one place, use it for learning algorithms.
• XGBoost — the de facto standard for gradient boosting trees, the champion tool for tabular data.
• LightGBM — Microsoft's faster gradient boosting, often outperforms earlier XGBoost at large scale.
• statsmodels — dedicated to statistical modeling and testing: p-values for linear regression, time series ARIMA, hypothesis testing.

Deep Learning

• PyTorch — the current mainstream deep learning framework in both research and industry: dynamic graph, largest ecosystem. See the site's Deep Learning Fundamentals and Optimization sections.
• TensorFlow / Keras — the veteran framework; Keras' high-level API is beginner-friendly, production deployment (TFLite/TF Serving) is mature.
• Hugging Face Transformers (library) — a unified interface for loading pre-trained models: BERT, GPT, LLaMA, T5 usable in one line, with a companion Trainer that simplifies fine-tuning.
• Hugging Face Datasets (library) — a dataset library paired with Transformers: lazy loading, memory mapping, sharding — processing large text datasets without blowing up memory.

The Golden Rule for Choosing Libraries

Use scikit-learn for learning principles, XGBoost/LightGBM for production, PyTorch for deep learning, Transformers for NLP. Don't use deep learning for tabular data, and don't use sklearn to train large models — tool selection itself is an engineering decision.

Data Engineering Tools

When data volume grows, "having data" isn't enough — you also need to "manage data". This toolchain turns data into a sustainably evolving asset; the full workflow is in Data and Data Engineering.

Tool	URL	Problem It Solves
DVC (Data Version Control)	dvc.org	Manage data the same way you manage code with Git: dataset versioning, rollback, reproducible experiments
Feast	feast.dev	Open-source feature store: training features and online inference features share the same definition,solving "training/online inconsistency"
Apache Airflow	airflow.apache.org	The most popular batch-processing workflow scheduler: data pipeline scheduling with timing/dependencies
Prefect	www.prefect.io	A more approachable modern data flow orchestrator, natively Python with strong dynamic scheduling
Dagster	dagster.io	An orchestration framework built around data assets, treating "data dependencies" as a first-class citizen
Great Expectations	greatexpectations.io	Data quality testing framework: declare "what data should look like", automatically validates after pipelines run
dbt	www.getdbt.com	Engineering tool for data transformations using SQL, the standard tool for analytics engineers
MLflow	mlflow.org	Unified entry point for experiment tracking + model registry + deployment, the first stop for MLOps adoption
DuckDB	duckdb.org	Embedded analytical database, run SQL directly on pandas/Parquet files — a single-machine big-data analysis tool

Don't Roll Out the Full Suite From Day One

For beginners and small-to-medium projects, DVC + MLflow + Great Expectations — a three-piece set covering the most critical pain points of "versioning, experiments, and quality" — is sufficient, and all are open-source with single-machine deployment. Hold off on Airflow/Feast until you truly have deployment and multi-team collaboration needs; introducing toolchains too early is an anti-pattern.

How to Choose Datasets for Practice

Choosing data is the first step of practice, and where many learners get stuck. Here are four pieces of experience.

1. Pick Difficulty by Learning Phase

• Beginner (first 3 projects): pick small, clean, well-documented ones — Iris, Titanic, MNIST. The goal is to run through the full workflow, not to chase scores.
• Advanced (after mastering fundamentals): switch to data with noise, missing values, and real-world pain points — House Prices (79 features teaching you feature engineering), Credit Card Fraud (highly imbalanced teaching you to change metrics).
• Further advanced: dig into your chosen direction — CV: CIFAR-10 → ImageNet subset; NLP: IMDB → SQuAD; Recommendation: MovieLens; Time series: Air Passengers → ETTh. See Learning Paths for route planning.

2. Pick Tasks by Practice Goal

Skill to Practice	Recommended Data	Why
Feature engineering	House Prices, Titanic	Many missing values, categorical features, long-tail distributions — the most operations to perform
Model tuning	CIFAR-10, IMDB	Massive community solution write-ups, easy to compare your hyperparameter choices
Deep learning	MNIST → Fashion-MNIST → CIFAR-10	Smooth difficulty progression, each level adding "real problems" (e.g., data augmentation)
Engineering deployment	Any real data + your own data pipeline	The focus is not on model accuracy but on DVC versioning, reproducibility, and monitorability
Interview portfolio piece	Pick one where you can tell a "business story"	See Portfolio Projects and Build Your Own

3. Check Whether the Data Is "Worth Your Time"

• Licensing: Kaggle competition data has clear competition agreements; UCI/OpenML are mostly permissive. Always confirm licensing terms before commercial or public use.
• Source and timeliness: Pre-2010 datasets are fine for teaching but not for discussing "current state". Be wary of data with questionable annotation quality (e.g., scraped data).
• Is there still "something to do": If a dataset has already been squeezed to 99%+ accuracy, don't expect it to produce results — go to Papers with Code for harder benchmarks.
• Are you genuinely interested: Willingness to invest another 20 hours into "this topic" is the greatest driver for completing a project.

4. Build Your Own "Data Garden"

Keep notes on good datasets you've used, pitfalls you've encountered (which dataset was dirty, which had misannotations), and verified loading code. This habit, paired with the Awesome Resources and Glossary, will make your learning accumulation grow thicker and thicker.

Further Reading

• Data and Data Engineering — complete methodology for data leakage, cleaning, and versioning
• Feature Engineering — transforming raw data into features models can consume
• Glossary — accurate definitions for all terms appearing on this page
• Awesome Resources — more books, courses, and papers
• Learning Paths — route planning from beginner to advanced
• Build Your Own and Portfolio Projects — how to turn practice into showcaseable results

All links on this page point to official sources. Dataset scale information is current as described on each official page; licensing terms should be verified before downloading.