[Jinja] GPT4All 3.5.1 breaks GPT4All-Community/gemma-2-9b-it-GGUF

[ThiloteE]

Bug Report

1. System message field triggers following error: "System message is not "plain text""
2. Jinja default chat template triggers following error: "Syntax error: 4:2: error: Unexpected token '}'. Expected: '>' '}}{% endif %} -^-------"

Default Jinja template in GPT4All (broken)

{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '
' + message['content'] | trim + '<end_of_turn>
' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model
'}}{% endif %}

Jinja template that doesn't trigger errors (still broken!):

{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '
' + message['content'] | trim + '<end_of_turn>
' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model
' }}{% endif %}

Steps to Reproduce

Download and use the model

Expected Behavior

1. Model just works.
2. gemma-2-9b-it was trained without system prompt, so having none should just work instead of triggering an error.
3. Default Jinja template works out of the box.

Your Environment

• GPT4All version: 3.5.1
• Operating System: Windows 10
• Chat model used (if applicable): https://huggingface.co/GPT4All-Community/gemma-2-9b-it-GGUF

[ThiloteE]

Oh, once the jinja template is "fixed" (still broken), the error message related to the system message also vanishes.

[ThiloteE]

Working/fixed template:

{{ bos_token }}
{%- if messages[0]['role'] == 'system' %}
	 {{- raise_exception('System role not supported') }}
{%- endif %}

{%- for message in messages %}
	 {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}
		  {{- raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}
	 {%- endif %}
	 {%- if (message['role'] == 'assistant') %}
		  {%- set role = 'model' %}
		  {%- else %}{% set role = message['role'] %}
	 {%- endif %}
	 {{- '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}
{%- endfor %}

{%- if add_generation_prompt %}
	 {{- '<start_of_turn>model\n' }}
{%- endif %}

[tcfurrer]

With that template on 3.5.3 I'm getting:

Error: Failed to parse chat template: 1:1: error: Unexpected exception occurred during template processing.
Exception: Jinja template error: Conversation roles must alternate user/assistant/user/assistant/...

That's happening when the System Message is just "You are a helpful assistant."

In case it matters, I'm using Mistral-Small-Instruct-2409-Q5_K_M.gguf

[ThiloteE]

This template is for the gemma model and based on its original template. The gemma model does not support a system prompt. If you wanted to add a system prompt, the template would have to be different. I get the same exception, as soon as I add a system message in the model settings. If you remove the system prompt, it also should work to a limited degree for the mistral model, but I recommend to come up with a proper template for Mistral based on its original template.