Thinking in LM Studio
It doesn't seem to work well at all with temp 0.6, min_p 0.01, rep_pen OFF in LM Studio. Also, it's thinking is not hidden. What are correct settings?
If you're seeing poor results with temp=0.6 and min_p=0.01, I recommend switching to:
temperature=1.0
top_p=0.95
This tends to yield more fluent and natural outputs.
Regarding the model showing its “thinking”: this is intentional.
Our model is designed to show its full reasoning process, as it's primarily optimized for step-by-step reasoning and chain-of-thought quality.
The model does not start its reasoning with <think> tag, and as a result the reasoning does not collapse to a "Thought process" block in ui
Ok with these settings this model still seems to think muuuch much longer than cogito 32b... And my hardware is too slow to wait that long. Possibly lm studio needs an update...
The model does not start its reasoning with
<think>tag
Actually it does. But because it's part of the input(part of chat template), it doesn't show on the output.
Current solution might be involving to
- remove those token from the chat template (the model might not think)
- use another ui that support text completion (eg. SillyTavern, Kobold) (better use this)