[BFCL] Gemma Prompt Structure

[mitchelldehaven]

Describe the issue
The prompt structure for Gemma 3's leaderboard placement is:

class GemmaHandler(OSSHandler):
    def __init__(self, model_name, temperature) -> None:
        super().__init__(model_name, temperature)

    @override
    def _format_prompt(self, messages, function):
        """
        "bos_token": "<bos>",
        "chat_template": "{{ bos_token }}\n{%- if messages[0]['role'] == 'system' -%}\n    {%- if messages[0]['content'] is string -%}\n        {%- set first_user_prefix = messages[0]['content'] + '\n\n' -%}\n    {%- else -%}\n        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '\n\n' -%}\n    {%- endif -%}\n    {%- set loop_messages = messages[1:] -%}\n{%- else -%}\n    {%- set first_user_prefix = \"\" -%}\n    {%- set loop_messages = messages -%}\n{%- endif -%}\n{%- for message in loop_messages -%}\n    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}\n        {{ raise_exception(\"Conversation roles must alternate user/assistant/user/assistant/...\") }}\n    {%- endif -%}\n    {%- if (message['role'] == 'assistant') -%}\n        {%- set role = \"model\" -%}\n    {%- else -%}\n        {%- set role = message['role'] -%}\n    {%- endif -%}\n    {{ '<start_of_turn>' + role + '\n' + (first_user_prefix if loop.first else \"\") }}\n    {%- if message['content'] is string -%}\n        {{ message['content'] | trim }}\n    {%- elif message['content'] is iterable -%}\n        {%- for item in message['content'] -%}\n            {%- if item['type'] == 'image' -%}\n                {{ '<start_of_image>' }}\n            {%- elif item['type'] == 'text' -%}\n                {{ item['text'] | trim }}\n            {%- endif -%}\n        {%- endfor -%}\n    {%- else -%}\n        {{ raise_exception(\"Invalid content type\") }}\n    {%- endif -%}\n    {{ '<end_of_turn>\n' }}\n{%- endfor -%}\n{%- if add_generation_prompt -%}\n    {{'<start_of_turn>model\n'}}\n{%- endif -%}\n",
        """
        formatted_prompt = "<bos>"

        if messages[0]["role"] == "system":
            first_user_prefix = messages[0]["content"].strip() + "\n\n"
            messages = messages[1:]
        else:
            first_user_prefix = ""

        is_first = True
        for message in messages:
            formatted_prompt += f"<start_of_turn>{message['role']}\n{first_user_prefix if is_first else ''}{message['content'].strip()}<end_of_turn>\n"
            is_first = False

        formatted_prompt += f"<start_of_turn>model\n"

        return formatted_prompt

I've been using Gemma and trying to determine the correct prompt structure. According to Google's documentation and other places (Phil Schmid's blog https://www.philschmid.de/gemma-function-calling), the Gemma 3 model doesn't actually use a system prompt role. Rather, what would be system instructions are placed under the "user" role in the first message.

However, Huggingface's documentation for Gemma 3 users a similar implementation to what is used here.

I'm trying to determine if the Gemma tool prompt is correctly formatted here.

[HuanzhiMao]

Hey @mitchelldehaven ,

Great question.

Gemma does let you pass a system-role message, but during its Jinja2 chat template rendering, that text is inserted under a user tag. That’s exactly how the Hugging Face template works, and BFCL mirrors the same behavior.

So, although you can supply a system message, once the prompt is compiled, Gemma only sees the first turn as coming from the user role, followed by alternating user/model turns.

Hope that clears things up!