LlamaFactory Prompt Templates: Complete Guide to 80+ Built-In Chat Formats

LlamaFactory supports over 80 built-in prompt templates including Alpaca, LLaMA-2, LLaMA-3, ChatML, Qwen, and Gemma, all registered in src/llamafactory/data/template.py and selectable via CLI, YAML, Web UI, or Python API.

LlamaFactory is a unified open-source framework for fine-tuning large language models. Understanding the prompt templates supported by LlamaFactory is essential because these templates define how multi-turn conversations are serialized into the token sequences that models actually process during training and inference.

What Are LlamaFactory Prompt Templates?

A prompt template in LlamaFactory is a Python class that encodes conversation history—system messages, user turns, and assistant responses—into a formatted string compatible with a specific model family.

In src/llamafactory/data/template.py, each template is registered via the register_template function, which stores template objects in the global dictionary TEMPLATES. The string name passed to register_template becomes the template identifier used throughout the framework.

# From src/llamafactory/data/template.py

@register_template(name="llama3")
def get_llama3_template() -> Template:
    return Template(
        format_user=StringFormatter(slots=["<|start_header_id|>user<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
        format_assistant=StringFormatter(slots=["<|start_header_id|>assistant<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
        format_system=StringFormatter(slots=["<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
        stop_words=["<|eot_id|>"],
    )

Complete List of Supported Prompt Templates

LlamaFactory organizes its 80+ templates by model family and use case. Below are the most commonly used identifiers, with the full registry available in src/llamafactory/data/template.py.

Classic Instruction Formats

These templates follow the original "Instruction / Response" paradigm popularized by early fine-tuning datasets.

• alpaca – The classic Alpaca format with ### Instruction: and ### Response: blocks (line 653 in template.py)

• vicuna – Vicuna-style USER: and ASSISTANT: turns with optional system messages (line 1523)

• default – Minimal fallback format using Human: and Assistant: prefixes (line 834)

Modern Chat Formats

These templates implement the specific token conventions required by contemporary foundation models.

• llama2 – Meta LLaMA-2 format with [INST] and [/INST] tokens, including system message injection via <<SYS>> (line 1313)
• llama3 – LLaMA-3 format with <|start_header_id|> and <|eot_id|> tokens, supporting reasoning tags (line 1331)
• mistral_small – Mistral Small format using [INST] … [/INST] with efficient EOS handling (line 1702)
• gemma – Google Gemma format with <start_of_turn> and <end_of_turn> tags (line 937)

Multilingual and Specialized Templates

Templates optimized for specific languages or model families with unique tokenization requirements.

• chatml – HuggingFace ChatML format using <|im_start|>user and <|im_end|> tokens

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}