README.md

Jinja Templates

A guide to Jinja2 templates for LLM chat templates.

Contents

• Glossary — key terms and definitions
• Basics — syntax, variables, filters
• Built-in Filters — filter reference

What is a Chat Template?

A chat template is a Jinja2 template that transforms structured chat messages into a text format understood by a specific LLM.

Input data:

{
  "messages": [
    {"role": "system", "content": "You are helpful assistant"},
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi there!"}
  ]
}

Output text (ChatML format):

<|im_start|>system
You are helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there!<|im_end|>

Why is this needed?

Different models are trained on different formats:

• ChatML — <|im_start|>role\ncontent<|im_end|>
• Llama — [INST] content [/INST]
• Mistral — <s>[INST] content [/INST]

Chat templates abstract these differences.

Our Template

Main project template: docker/llama/config/templates/qwen/default.jinja

Features:

• Multimodal — image and video support
• Tool calling — function call format
• Thinking — <think> blocks for chain-of-thought