Synthetic Data, Carefully

Name: Synthetic Data, Carefully
Availability: InStock

When Generated Data Helps, and When It Teaches Models Their Own Mistakes

Subscribe to get the complete book. Your PDF downloads immediately.

01/Overview

What's in the book

Synthetic data scales until the model starts learning its own mistakes. Where the technique helps and where it quietly rots.

TrainingResearch

02/Inside the PDF

PDF page

FMFront Matter: Synthetic Data, Carefully—
INTIntroduction: The Sentence That Should Make You Nervous15
01The Polished-Ticket Problem20
02A Taxonomy of Generated Data: and Why We Reach for It27
03Generated Data Without Lineage Is Operational Debt34
04Manifests, Data Cards, and the Diff That Regenerates a Dataset41
05Patterns That Earn Their Keep49
06Diversity Is a Measurement, Not a Vibe57
07The Gates After Generation65
08Judges Disagree, and That Is the Useful Part72
09Evaluation Is Where Synthetic Data Does the Most Damage80
10Model Collapse, Honestly87
11Mixture Ratios, Anchors, and Post-Training Monitoring94
12The Privacy Myth and What Generated Data Actually Leaks101
13Poisoning, Approvals, and the Governance of Generation108
14Playbooks I: Classification, Extraction, and Safety Red-Teaming115
15Playbooks II: RAG Evaluation, Fine-Tuning Tone, Low-Resource, Code, and Cautious Domains122
16Operating Synthetic Data with CAREFUL130
AAppendix A: Back Matter137

The full chapters, illustrations, and reference material are included in the downloadable PDF.