Implementation:Unslothai Unsloth Get Chat Template

Knowledge Sources	Unsloth HuggingFace Chat Templating
Domains	NLP, Data_Preprocessing, Tokenization
Last Updated	2026-02-07 00:00 GMT

Overview

Concrete tool for applying chat templates to HuggingFace tokenizers provided by the Unsloth library.

Description

The get_chat_template function configures a HuggingFace tokenizer with the appropriate Jinja2 chat template for a given model family. It supports 40+ model families (Llama 3, Mistral, ChatML, Gemma, Qwen, Phi, etc.) via the internal CHAT_TEMPLATES registry. The function also handles EOS token mapping, system message injection, and role/content field remapping for ShareGPT-format datasets.

Usage

Import this function immediately after loading a tokenizer and before formatting training data. Use it when fine-tuning any chat or instruction model to ensure the tokenizer applies the correct conversational template.

Code Reference

Source Location

• Repository: unsloth
• File: unsloth/chat_templates.py
• Lines: L2123-2350

Signature

def get_chat_template(
    tokenizer,
    chat_template = "chatml",
    mapping = {"role": "role", "content": "content", "user": "user", "assistant": "assistant"},
    map_eos_token = True,
    system_message = None,
):
    """
    Applies a chat template to the tokenizer and configures EOS token mapping.

    Args:
        tokenizer: HuggingFace PreTrainedTokenizer to configure.
        chat_template (str): Template name (e.g., "chatml", "llama-3", "mistral",
            "gemma", "qwen-2.5", "phi-4") or a custom template tuple.
        mapping (dict): Role/content field mapping for ShareGPT format datasets.
            Keys: "role", "content", "user", "assistant".
        map_eos_token (bool): Whether to map the EOS token to the template's
            stop token. Default True.
        system_message (str): Optional system prompt to inject into the template.

    Returns:
        Modified tokenizer with Jinja2 chat_template set and EOS token configured.
    """

Import

from unsloth.chat_templates import get_chat_template

I/O Contract

Inputs

Name	Type	Required	Description
tokenizer	PreTrainedTokenizer	Yes	HuggingFace tokenizer to configure
chat_template	str	No	Template name from CHAT_TEMPLATES registry (default: "chatml")
mapping	dict	No	Role/content field mapping for ShareGPT datasets
map_eos_token	bool	No	Map EOS token to template stop word (default: True)
system_message	str	No	Optional system prompt injected into template

Outputs

Name	Type	Description
tokenizer	PreTrainedTokenizer	Modified tokenizer with chat_template Jinja2 string and EOS token configured

Usage Examples

Basic Chat Template Application

from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template

# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Llama-3.2-1B-Instruct",
    max_seq_length=2048,
    load_in_4bit=True,
)

# Apply Llama 3 chat template
tokenizer = get_chat_template(
    tokenizer,
    chat_template="llama-3",
)

# Format a conversation
messages = [
    {"role": "user", "content": "What is machine learning?"},
    {"role": "assistant", "content": "Machine learning is a branch of AI..."},
]
formatted = tokenizer.apply_chat_template(messages, tokenize=False)

With System Message and ShareGPT Mapping

from unsloth.chat_templates import get_chat_template

tokenizer = get_chat_template(
    tokenizer,
    chat_template="chatml",
    mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},
    system_message="You are a helpful assistant.",
)

Related Pages

Implements Principle

• Principle:Unslothai_Unsloth_Chat_Template_Configuration

Requires Environment

• Environment:Unslothai_Unsloth_Python_Transformers