Phase 6 · Modern LLM engineering · Post-training
Instruction and preference data
The datasets that shape assistant behavior.
Open this lesson Create a free account
Why it matters
A model becomes the assistant its post-training data describes. The examples people write, and the choices they make between two answers, decide its tone, its length, what it refuses and what it flatters. If you fine-tune a model, run evals with an LLM judge, or just want to understand why assistants behave as they do, the data and its quirks are where to look.
What you'll be able to do
- Describe the three main kinds of post-training data and how each is collected
- Turn pairwise preferences into scores with the Bradley-Terry model
- Measure annotator agreement and detect position bias in judgments
- Recognize the biases that preference data passes on to models
Before this lesson
Next in this phase: RLHF