Free account
Read this concept with a free account
Read this concept at no cost by signing in with Google.
no card needed
Core
Synthetic Data and Distillation: Training Smaller Models on Generated Examples
Distillation trains a smaller model to imitate a larger one, often on synthetic examples the larger model generates. It can cut serving cost and latency sharply for a narrow task. It also inherits the teacher's errors, narrows variety and can breach provider terms. When to use it, how to check the data and how to prove the student is good enough.
CONCEPT / MODELS & PLATFORMS
RELATED CONCEPTS
PRACTISE THIS CONCEPT
AI Platform & Model PMThe team wants to train our smaller model on synthetic data generated by a frontier provider's model. Do you approve it?→LLM & RAG FundamentalsWhat is model distillation, and when would you, as a PM, ask the team to do it?→LLM & RAG FundamentalsShould our phone's AI assistant run models on the device, in the cloud, or both? Make the strategic case.→Safety, Security & GovernanceThe ML team wants to use customer conversations to improve our models. What do you need to decide before saying yes?→AI System DesignAn AI reply-suggestion feature in our messaging app will get a billion requests a day. How do you make cost and latency work?→AI UX & Human-in-the-LoopLeadership wants to start training our models on consumer chat conversations. Design how users are asked.→