The <think> tags
Hello,
The model isn’t picking up the think tags. Will this start working after a llama.cpp update? Currently, the reasoning mode doesn’t work by default. I got it working by following advice from https://huggingface.co/ilintar/THUDM_GLM-Z1-9B-0414_iGGUF — I copied the Jinja template from there.
Hello,
The model isn’t picking up the think tags. Will this start working after a llama.cpp update? Currently, the reasoning mode doesn’t work by default. I got it working by following advice from https://huggingface.co/ilintar/THUDM_GLM-Z1-9B-0414_iGGUF—I copied the Jinja template from there.
I just copied prompt template from GLM-4-32B in LM Studio and after that tag "think" works correctly now
Hello,
The model isn’t picking up the think tags. Will this start working after a llama.cpp update? Currently, the reasoning mode doesn’t work by default. I got it working by following advice from https://huggingface.co/ilintar/THUDM_GLM-Z1-9B-0414_iGGUF—I copied the Jinja template from there.
I just copied prompt template from GLM-4-32B in LM Studio and after that tag "think" works correctly now
Thanks, that helped! By the way, with the temperature set at 0.6 (as recommended), the responses sometimes lack detail. At 0.7 or 0.8, the answers are much more complete and logical. This applies to chat interactions, not programming.
Hello,
The model isn’t picking up the think tags. Will this start working after a llama.cpp update? Currently, the reasoning mode doesn’t work by default. I got it working by following advice from https://huggingface.co/ilintar/THUDM_GLM-Z1-9B-0414_iGGUF—I copied the Jinja template from there.
I just copied prompt template from GLM-4-32B in LM Studio and after that tag "think" works correctly now
Thanks, that helped! By the way, with the temperature set at 0.6 (as recommended), the responses sometimes lack detail. At 0.7 or 0.8, the answers are much more complete and logical. This applies to chat interactions, not programming.
I've tested GLM-4 and GLM-Z1 for programming, my opinion that GLM-4 gives much more better results any way... Really strange
Hello,
The model isn’t picking up the think tags. Will this start working after a llama.cpp update? Currently, the reasoning mode doesn’t work by default. I got it working by following advice from https://huggingface.co/ilintar/THUDM_GLM-Z1-9B-0414_iGGUF—I copied the Jinja template from there.
I just copied prompt template from GLM-4-32B in LM Studio and after that tag "think" works correctly now
Thanks, that helped! By the way, with the temperature set at 0.6 (as recommended), the responses sometimes lack detail. At 0.7 or 0.8, the answers are much more complete and logical. This applies to chat interactions, not programming.
I've tested GLM-4 and GLM-Z1 for programming, my opinion that GLM-4 gives much more better results any way... Really strange
They're suited for different tasks. For example, I had a list and asked to sort it—the standard GLM handled everything perfectly (didn’t miss a single line), which was great. Then, I gave the same task to Z1: it didn’t process all lines but instead described the sorting structure in detail, using at most 50 out of 300 lines, while thoroughly explaining the logic, steps, and even notes on why. Now, here’s the catch: if we take that same logic and ask the standard model to sort accordingly, it succeeds. But if we ask Z1 to apply the exact same logic, it still skips some lines.