-
Notifications
You must be signed in to change notification settings - Fork 144
Open
Description
I've been using tinker to finetune qwen3 30b a3b instruct on conversations involving tool use, and just wanted to bring up a discrepancy in the qwen3 renderer in the cookbook vs. the jinja templates on huggingface.
In the jinja template, there are the lines
{%- elif message.role == "tool" %}
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
{{- '<|im_start|>user' }}
{%- endif %}
which render tool responses with the user role, while in the cookbook the role is used as-is,
ob_str = f"{maybe_newline}<|im_start|>{message['role']}\n"
in renderers.py, lines 332 and 445.
The rendering of elements under the "tools" key of a conversation as part of the system prompt is also not implemented, but that seems to be the case for all renderers, not just the qwen ones
Metadata
Metadata
Assignees
Labels
No labels