Synthetic data
Synthetic data is data generated artificially rather than collected directly from the target process, often to approximate its properties or meet specified requirements.
Outlier Detection in Python
Artificially generated data designed to have controlled properties rather than being collected from real-world observations.
- How it is used
- Uses generated records to understand outlier detectors, tune and test them, or train detectors that distinguish real from synthetic data.
- How deep it goes
- Works through feature-wise generation, joint-pattern modeling, doping, and evaluation, while noting that simulations require more setup.
- Examples
- Uses two-dimensional clusters, breast-cancer records, and baseball features to demonstrate detector scores, real-versus-fake classification, and lost feature relationships.
- What you can do after
- Build it from scratch with feature-wise sampling, then evaluate whether generated cases receive higher outlier scores than real records.
Read how Outlier Detection in Python explains synthetic data →
Causal Inference for Data Science
Synthetic data are artificially generated observations produced from a specified model rather than collected from real-world events.
- How it is used
- Tests causal methods against known data-generating processes before applying them to real-world datasets.
- How deep it goes
- Works through variable-generation equations, noise, potential outcomes, and comparisons between naive and causal estimates.
- Examples
- Uses population-driven arcade revenue and doctorates, hotel pricing, game levels, instruments, and a movie-release time series.
- What you can do after
- Evaluate/understand whether an estimator recovers a known causal effect and how confounding can bias naive analysis.
Read how Causal Inference for Data Science explains synthetic data →
Data for All
Simulated data is data generated by modeling or simulating conditions, processes, or outcomes rather than directly observing them.
- How it is used
- Uses synthetic data to explore decisions where the activity of interest has not yet been observed.
- How deep it goes
- Explains how observed data informs generated figures and distinguishes scenario testing, without providing an implementation procedure.
- Examples
- Works through proposed retail stores and the figures used to compare markets, locations, pricing, and staffing.
- What you can do after
- Evaluate/understand whether planning figures describe observed activity or generated possibilities.
Introduction to Generative AI, Second Edition
A synthetic dataset is a dataset whose records are artificially generated rather than directly collected from real-world subjects or events.
- How it is used
- Supports privacy-preserving research, dataset augmentation, rare-scenario simulation, and scalable evaluation.
- How deep it goes
- Explains benefits and limitations through applications, but stops short of generation methods or technical implementation.
- Examples
- Uses synthetic genome data for healthcare and generated prompts for large language model evaluation.
- What you can do after
- Evaluate/understand when synthetic data can extend authentic datasets and where realism, distribution, privacy, or coverage may fail.
Read how Introduction to Generative AI, Second Edition explains synthetic data →
Data Without Labels
Synthetic data is artificially generated data designed to reproduce selected characteristics or distributions of real data.
- How it is used
- Augments scarce or small datasets for machine-learning training and supports generative applications across healthcare and other business domains.
- How deep it goes
- Explains dataset augmentation, GAN components, adversarial losses, and autoencoder generation, while stopping short of implementation code.
- Examples
- Uses healthcare data augmentation, random-noise generation, GAN real-versus-synthetic classification, and autoencoder reconstruction.
- What you can do after
- Apply it via GANs or autoencoders to expand training data and reduce overfitting risk.
Software Testing with Generative AI
Synthetic data is artificially generated data designed to reproduce relevant characteristics of data from a real-world domain.
- How it is used
- Supports AI training, fine-tuning, and rapid test-data generation when suitable real data is difficult to prepare.
- How deep it goes
- Defines the technique, contrasts it with real data, and identifies cost and quality risks without detailing generation algorithms.
- Examples
- Uses Alpaca’s synthetic instruction records and generated test-data examples to illustrate practical dataset construction.
- What you can do after
- Apply synthetic data to supply training, fine-tuning, or testing examples while checking generation cost and realism.
Read how Software Testing with Generative AI explains synthetic data →
Learn AI-Assisted Python Programming, Second Edition
- How it is used
- Uses controllable datasets to test functions against known answers and difficult cases without relying on rare real examples.
- How deep it goes
- Explains dataset design and expected outputs, but stops short of describing generation algorithms or statistical fidelity.
- Examples
- Uses an artificial NFL statistics file containing repeated players, non-quarterbacks, zero values, and negative passing yards.
- What you can do after
- Diagnose/debug functions by testing known outputs against realistic records and intentionally included edge cases.
Read how Learn AI-Assisted Python Programming, Second Edition explains synthetic data →
Reinforcement Learning from Human Feedback
Synthetic data is data generated by an algorithmic or model-based process rather than directly collected from the target phenomenon.
- How it is used
- Supplies scalable prompts, completions, feedback, and filtering signals across post-training and RLHF pipelines.
- How deep it goes
- Explains generation, filtering, collapse risks, and Constitutional AI construction, while deferring detailed model-training objectives.
- Examples
- Uses Constitutional AI revisions and preference judgments plus Tülu 3’s million-scale synthetic instruction and preference stages.
- What you can do after
- Apply it via language-model generation, judging, filtering, and preference-data pipelines for post-training.
Read how Reinforcement Learning from Human Feedback explains synthetic data →
Machine Learning System Design
Artificially generated data designed to approximate the characteristics of data from a real-world domain.
- How it is used
- Uses generated examples to bootstrap ML projects when suitable real-world data is scarce.
- How deep it goes
- Defines major sources and explains realism tradeoffs, but stops short of generation algorithms or implementation details.
- Examples
- Uses scientific simulators, rendered environments, generative AI, and augmented-reality recordings with estimated lighting.
- What you can do after
- Evaluate whether generated data is realistic and domain-relevant enough to bootstrap model development.
Read how Machine Learning System Design explains synthetic data →
Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI
Synthetic data is artificially generated data designed to resemble the characteristics or distribution of real data without representing actual observations.
- How it is used
- Expands existing tabular datasets for machine-learning and deep-learning workflows.
- How deep it goes
- Explains benefits and introduces GANs, VAEs, TVAE generation, and adversarial validation without deriving model internals.
- Examples
- Uses German credit data, SDV metadata, a TVAE synthesizer, and random-forest adversarial validation.
- What you can do after
- Build synthetic tabular data with SDV and evaluate distinguishability through adversarial validation.
Read how Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI explains synthetic data →
The AI Pocket Book
Artificially generated training data used as a substitute for or augmentation to real data.
- How it is used
- Frames synthetic data as a possible alternative to painstaking collection and manual labeling, with narrow practical successes.
- How deep it goes
- Explains the labeling requirement and contrasts open-ended tasks with self-play, but stops short of a broader generation workflow.
- Examples
- Uses AlphaZero’s self-played Go games, LLM next-token prediction, and image-category labeling as contrasting examples.
- What you can do after
- Evaluate when simulation can produce useful labeled training data and recognize why it fails for many open-ended applications.
AI-Powered Search
Synthetic training data is artificially generated data created to support the training or evaluation of machine-learning models.
- How it is used
- Generates synthetic queries and relevance signals to support search ranking, click modeling, and other data-hungry relevance techniques.
- How deep it goes
- Demonstrates prompting, multi-document context, generated judgments, and parsing, but does not detail foundation-model training internals.
- Examples
- Uses an outdoors collection, minimalist-shoe queries, retrieved result sets, relevance indexes, and parsed query-document records.
- What you can do after
- Apply it via an LLM prompt and parser to create synthetic query-document relevance pairs for search training.
Building LLM Applications with DSPy
Artificially generated data designed to resemble data from a real domain or process.
- How it is used
- Builds or expands evaluation sets for LM-based classifiers when labeled real-world examples are scarce or incomplete.
- How deep it goes
- Explains two generation workflows and review concerns, while warning that generated examples may be unrealistic or unusable.
- Examples
- Generates five intent-specific queries or two alternative phrasings per labeled example.
- What you can do after
- Apply DSPy’s Predict module to generate intent examples or label-preserving variations for an evaluation set.
Read how Building LLM Applications with DSPy explains synthetic data →
Machine Learning Platform Engineering
Synthetic data is artificially generated data designed to resemble observations from a target domain.
- How it is used
- Supports ML data collection when real examples are difficult to obtain or expensive to label.
- How deep it goes
- Defines procedural generation and labeling benefits but stops short of explaining generation algorithms or fidelity evaluation.
- Examples
- Uses identity-document images and potential fraud examples as concrete synthetic-data applications.
- What you can do after
- Build labeled computer-vision datasets from procedurally generated images while reducing manual annotation.
Read how Machine Learning Platform Engineering explains synthetic data →
Financial AI in Practice
The creation of artificial data that reproduces selected properties of real data without directly using the original records.
- How it is used
- Supports testing credit-risk systems in non-production environments subject to regulatory constraints.
- How deep it goes
- Flags synthetic data generation as necessary but does not describe generation methods or validation.
- Examples
- Warns that randomly generated Luhn-compliant credit card numbers can still cause regulatory violations.
Related concepts
- generative modelingproduced by · Generative modeling is a machine-learning approach that learns patterns or a data distribution to generate new samples resembling the data.
- datasetkind of · A dataset is an organized collection of related data items, records, or observations treated as a unit for storage, processing, analysis, or computation.
- language modelproduced by · A language model is a computational model that represents patterns in linguistic tokens to estimate sequence probabilities and predict or generate language from context.
- simulationproduced by · A simulation is a modeled or controlled enactment of a system, process, or scenario used to examine or reproduce behavior under specified conditions.
- Ragasproduced by · Ragas is an open-source framework and Python library for evaluating retrieval-augmented generation systems and applications.
- production datacontrasts with · Production data is data generated or collected through actual operational processes or real-world activity, rather than solely through artificial or controlled experimental settings.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.