Training & Model Adaptation · Glossary

Topic Glossary

Plain-English definitions for the concepts in this module, including the new Synthetic Data topic.

Term	Simple English	PM-friendly meaning
Training Data	Examples a model learns from.	The material that shapes learned patterns and behavior.
Synthetic Data	Data generated by rules, simulation, or an AI model.	A scalable, controlled input for training, evaluation, or testing—if quality is checked.
Pre-training	Broad initial training on large amounts of data.	The stage that creates general capabilities in a base model.
Fine-tuning	Additional training on focused examples.	A way to adapt a pre-trained model to a more specific task or behavior.
Instruction Tuning	Training on instructions paired with desired responses.	Helps a model follow task instructions more consistently.
Distillation	Using a larger or stronger model to teach a smaller model.	A way to transfer useful behavior into a cheaper or faster model.
Data Augmentation	Creating variations of existing examples.	Expands coverage while trying to preserve the original task meaning.
Evaluation Dataset	A set of cases used to measure performance.	Evidence for deciding whether a model or system is good enough.
Model Collapse	Loss of useful variation after repeated training on generated data.	A risk when synthetic outputs replace too much real-world diversity.

Synthetic Data is related to training data, but the terms are not identical: synthetic data describes where examples came from; training data describes how examples are used.