models

LiteRT-LM Chat Templates

This directory contains canonical prompt templates, metadata configurations, and schema definitions for models supported by LiteRT-LM.

LiteRT-LM uses Jinja2 (rendered hermetically via Minijinja) to transform structured conversation turns and tool declarations into model-specific prompt strings.

For the formal JSON Schema specification of all input variables, see chat_template_input_schema.json.

Input Variables

When rendering a prompt template, LiteRT-LM passes a context dictionary containing the following top-level fields:

Field	Type	Description
`messages`	`array`	(Required) The conversation history as a list of message objects.
`tools`	`array`	(Optional) Available tools (function declarations) that the model can invoke.
`enable_thinking`	`boolean`	(Optional) Whether the model should produce reasoning thoughts before answering.
`add_generation_prompt`	`boolean`	(Optional) Whether to append the model turn prefix to prompt generation (defaults to `true`).
`bos_token`	`string`	(Optional) Beginning-of-sequence token (e.g. `<bos>`, `<s>`).

Message Structure

Each entry in messages represents a single turn in the dialogue.

Roles

The role string specifies the author of the turn:

"system": System instructions, personality, guidelines, and tool schemas.
"user": Prompts and queries from the end user.
"assistant": Responses generated by the model (text and/or tool calls).
"tool": Outputs and return values from executed tools/functions.

Content Types

The content field is strictly a list of multimodal parts. Plain text prompts are wrapped in a text part object:

{
  "role": "user",
  "content": [
    {"type": "text", "text": "What is the capital of France?"}
  ]
}

Multimodal prompts with images, audio, or video include the corresponding part objects alongside text parts:

{
  "role": "user",
  "content": [
    {"type": "text", "text": "What is in this image?"},
    {"type": "image"}
  ]
}

Tool Calls

When the assistant requests one or more function invocations, the message includes a tool_calls list:

{
  "role": "assistant",
  "tool_calls": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": {"location": "San Francisco, CA"}
      }
    }
  ]
}

When multiple tool calls are invoked in parallel within a single turn:

{
  "role": "assistant",
  "tool_calls": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": {"location": "San Francisco, CA"}
      }
    },
    {
      "type": "function",
      "function": {
        "name": "get_time",
        "arguments": {"location": "San Francisco, CA"}
      }
    }
  ]
}

Note: arguments can be provided either as a JSON object or as a serialized JSON string. When tool_calls is present, content is optional and may be omitted or empty.

Tool Responses

When a function executes, its result is supplied as a message with role: "tool":

{
  "role": "tool",
  "content": [
    {
      "type": "tool_response",
      "name": "get_weather",
      "response": {"temperature": "18C", "condition": "Sunny"}
    }
  ]
}

When multiple tool calls are executed, their responses are bundled into a single message with multiple tool_response items in content:

{
  "role": "tool",
  "content": [
    {
      "type": "tool_response",
      "name": "get_weather",
      "response": {"temperature": "18C", "condition": "Sunny"}
    },
    {
      "type": "tool_response",
      "name": "get_time",
      "response": {"time": "14:30"}
    }
  ]
}

Tool Definitions

When tools are supplied, each item in the tools array follows the OpenAPI standard:

{
  "type": "function",
  "function": {
    "name": "get_weather",
    "description": "Fetch the current weather for a specified location.",
    "parameters": {
      "type": "object",
      "properties": {
        "location": {
          "type": "string",
          "description": "City and state, e.g. San Francisco, CA"
        },
        "unit": {
          "type": "string",
          "enum": ["celsius", "fahrenheit"]
        }
      },
      "required": ["location"]
    }
  }
}

Writing Templates for LiteRT-LM

Chat templates should be placed in the model's directory (e.g. models/<model_name>/chat_template.jinja).

Built-in Filters and Functions

The Minijinja engine provides several built-in utilities:

tojson: Serializes a value into formatted JSON string (e.g. {{ tool | tojson }}).
lstrip(chars) / rstrip(chars): Trims leading/trailing whitespace or specific characters.
strftime_now(format): Formats the current date/time into a string using the given format (e.g. {{ strftime_now("%d %b %Y") }}). Time-aware templates should use this function rather than referencing a raw now variable.
raise_exception(message): Raises an error during template rendering if an invariant is violated.
is none: Test for undefined or null variables.

Handling Tool Calls

Templates that support function calling typically check if tools is defined and loop through available function signatures:

{%- if tools is defined and tools -%}
<tools>
{%- for tool in tools %}
{{ tool.function | tojson }}
{%- endfor %}
</tools>
{%- endif %}

When rendering an assistant turn with function calls:

{%- if message.tool_calls is defined and message.tool_calls -%}
  {%- for tool_call in message.tool_calls %}
<tool_call>
{"name": "{{ tool_call.function.name }}", "arguments": {{ tool_call.function.arguments | tojson }}}
</tool_call>
  {%- endfor %}
{%- endif -%}

Handling Thinking Mode

If the model supports configurable thinking/reasoning modes, use the default filter:

{%- set thinking = enable_thinking | default(false) -%}
{%- if thinking -%}
<think>
</think>
{%- endif -%}

Examples

Standard Multi-turn Conversation

{
  "messages": [
    {
      "role": "system",
      "content": [
        {"type": "text", "text": "You are a concise, helpful assistant."}
      ]
    },
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "Hello!"}
      ]
    },
    {
      "role": "assistant",
      "content": [
        {"type": "text", "text": "Hi! How can I help you today?"}
      ]
    }
  ],
  "add_generation_prompt": true
}

Multimodal Inputs

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "image"},
        {"type": "text", "text": "Describe what is shown in this picture."}
      ]
    }
  ],
  "add_generation_prompt": true
}

Function Calling & Tool Response Turn

{
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "lookup_stock",
        "description": "Get current stock price for a ticker symbol.",
        "parameters": {
          "type": "object",
          "properties": {
            "symbol": {"type": "string", "description": "Stock ticker symbol, e.g. GOOG"}
          },
          "required": ["symbol"]
        }
      }
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What is the price of GOOG?"}
      ]
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "lookup_stock",
            "arguments": {"symbol": "GOOG"}
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": [
        {
          "type": "tool_response",
          "name": "lookup_stock",
          "response": {"price": "180.50", "currency": "USD"}
        }
      ]
    }
  ],
  "add_generation_prompt": true
}

Parallel Tool Calling Turn

{
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a location.",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {"type": "string", "description": "City name"}
          },
          "required": ["location"]
        }
      }
    }
  ],
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What is the weather in London and Paris?"}
      ]
    },
    {
      "role": "assistant",
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "get_weather",
            "arguments": {"location": "London"}
          }
        },
        {
          "type": "function",
          "function": {
            "name": "get_weather",
            "arguments": {"location": "Paris"}
          }
        }
      ]
    },
    {
      "role": "tool",
      "content": [
        {
          "type": "tool_response",
          "name": "get_weather",
          "response": {"location": "London", "temperature": "18C"}
        },
        {
          "type": "tool_response",
          "name": "get_weather",
          "response": {"location": "Paris", "temperature": "22C"}
        }
      ]
    }
  ],
  "add_generation_prompt": true
}

Testing and Validation

All chat template test cases are defined under testdata/input/*.json and validated using the chat_template_test Bazel macro from chat_template_test.bzl.

Rendered outputs are checked against golden reference files under testdata/golden/.

To add a test for a new model family, define in BUILD:

load("//models:chat_template_test.bzl", "chat_template_test")

chat_template_test(
    name = "chat_template_test",
    chat_template = "chat_template.jinja",
    golden_dir = "testdata/golden",
    input_dir = "testdata/input",
    pbtext_files = [":LlmMetadataProto.pbtext"],
)

To run all chat template tests across all models:

bazel test //models/...

Name		Name	Last commit message	Last commit date
parent directory ..
gemma4		gemma4
lfm2		lfm2
minicpm5		minicpm5
prem_1b_sql		prem_1b_sql
qwen2_5		qwen2_5
qwen3		qwen3
qwen3_5		qwen3_5
BUILD		BUILD
README.md		README.md
chat_template_input_schema.json		chat_template_input_schema.json
chat_template_test.bzl		chat_template_test.bzl
chat_template_test_runner.cc		chat_template_test_runner.cc

Navigation Menu

README.md

LiteRT-LM Chat Templates

Table of Contents

Input Variables

Message Structure

Roles

Content Types

Tool Calls

Tool Responses

Tool Definitions

Writing Templates for LiteRT-LM

Built-in Filters and Functions

Handling Tool Calls

Handling Thinking Mode

Examples

Standard Multi-turn Conversation

Multimodal Inputs

Function Calling & Tool Response Turn

Parallel Tool Calling Turn

Testing and Validation

FilesExpand file tree

models

Directory actions

More options

Directory actions

More options

Latest commit

History

models

Folders and files

parent directory

README.md

LiteRT-LM Chat Templates

Table of Contents

Input Variables

Message Structure

Roles

Content Types

Tool Calls

Tool Responses

Tool Definitions

Writing Templates for LiteRT-LM

Built-in Filters and Functions

Handling Tool Calls

Handling Thinking Mode

Examples

Standard Multi-turn Conversation

Multimodal Inputs

Function Calling & Tool Response Turn

Parallel Tool Calling Turn

Testing and Validation