Sage-A1.5 chat template: Verification Mode by reasoning effort
A modified chat_template.jinja of
InternScience/Sage-A1.5
that turns on its Verification Mode when the request asks for a high
reasoning effort, so that clients which know nothing about the
verification template argument can still use it.
GGUF quantizations of the model: na0x2c6/Sage-A1.5-GGUF.
Behavior
| Request | Verification Mode |
|---|---|
verification given (chat_template_kwargs) |
as given, same as the original template |
no verification, reasoning_effort is high, xhigh or max (case-insensitive) |
on, if there is a non-empty system message |
no verification, any other or no reasoning_effort |
off |
- An explicit
"verification": falsekeeps it off at any effort. - Without a system message, the effort leaves it off instead of raising an
error; an explicit
"verification": truestill raises one, as in the original. - To change the efforts, edit
verify_effortsin the template.
Everything else is unchanged; original.diff is the whole difference from
the upstream template.
Usage with llama-server
llama-server -m Sage-A1.5-Q4_K_M.gguf --jinja --chat-template-file chat_template.jinja
llama-server passes both of these to the template as reasoning_effort:
{"reasoning_effort": "high"}
{"chat_template_kwargs": {"reasoning_effort": "high"}}
The self-review is emitted as reasoning content, so read
reasoning_content as well as content.
Checked with llama.cpp (9b56ccf, /apply-template) and Python jinja2.
License
Apache-2.0, as the original model. Modified from the chat template of InternScience/Sage-A1.5.