Chat templates#

Chat templates turn structured messages into the exact prompt text expected by the selected tokenizer. AReno keeps this rendering close to rollout and serving so a training run and an OpenAI-compatible endpoint can use the same model conversation format.

Use chat templates when a model expects role-based messages such as system, user, assistant, or tool-related turns. Dataset loaders should keep raw task fields and normalized prompts separate; tokenizer-specific formatting belongs in the training or serving path.

Thinking mode#

Some reasoning or chat checkpoints expose an enable_thinking option through their tokenizer chat template. AReno’s CLI exposes --disable-thinking for training and serving, which passes enable_thinking=False when supported and falls back to the normal template call otherwise.

Where to go next#