Skip to main contentA Composition is the “recipe” for your fine-tune. Rather than just selecting a single dataset, a composition allows you to mix multiple data sources to find the optimal balance for your model.
Compositions as Recipes
Think of a composition as a way to combine various data sources:
- Mix Multiple Datasets: Combine real production data with synthetic data from Synthetic Conversation Trees.
- Balance Data Ratios: You might want 70% core tasks, 20% edge cases, and 10% “negative” examples (what NOT to do).
- Experimentation: Create multiple compositions with different dataset ratios to see which one yields the best performance during validation.
Next
