Chat Templates

Chat Templates are the bridge between a human-readable list of messages and the raw, model-specific token stream required by an LLM.

Conversational Roles

Standard chat templates use three primary roles:

• System: Persistent instructions that define the agent's identity, tone, and available tools.
• User: The human-provided query or instruction.
• Assistant: The model's reasoning, tool calls, and final responses.
• Tool: (Introduced in function-calling) The role used to feed the output of a tool back into the conversation.

Technical Implementation

Most modern frameworks (like transformers or smolagents) use Jinja2 templates to wrap messages with special tokens (e.g., <|im_start|>, [INST]) that the model was trained on. This ensures the model correctly identifies who said what.

References

• Source: 00_Raw/hf-agents-course-unit1.md
• hf-agents-course-moc
• function-calling