Turn-based training
Also called Multi-Turn Training
Training a model on whole back-and-forth conversations, not just single question-answer pairs, so it can hold a thread.
Think of it like
Rehearsing a full dialogue with an actor, not just isolated one-liners, so the scene actually flows.
Example
The model is fine-tuned on multi-turn chats where later replies must remember and build on what was said several turns earlier.
How it actually works
Single-turn tuning teaches a good one-shot answer but not conversational memory or consistency. Turn-based training uses multi-turn transcripts, typically with loss computed only on the assistant's turns, so the model learns to use prior context, stay coherent, and not contradict itself. It is what makes a model feel like it is following the conversation rather than answering each message from scratch.
For product teams
Why a good assistant remembers what you said three messages ago instead of resetting each time.
For engineers
Fine-tune on multi-turn transcripts with loss on assistant turns, teaching cross-turn context use and consistency.
Related
- Chat template — The format these transcripts use.
- Conversation — The multi-turn structure it trains on.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome