After training finishes, use the chat button on the Train block (or a run row in History) to talk to your new model:
- Modes: Adapter (fast — base + adapter), Merged (if a merge exists) and Base (the original model, for contrast).
- sampling: the chat UI replies with temperature 0.7 and up to 96 new tokens; both are precisely tunable through the API (
temperature,max_new). - the conversation is multi-turn — the model sees prior context.
- early stopping is normal: the model emits EOS when its answer is complete.
Tip: if replies turn to gibberish the learning rate was too high — see lesson 6. Low loss + on-style replies = a successful run.
