ORIGINAL REDDIT POST
Strategies for capping thinking on ds4 flash 0731
I like the outputs from this model, but DAMN does it over think. Has anyone found a robust fix for this that isn't just capping output tokens? Anyone working on a 'thinking cap' for it? Some combo of llama params, or (system?) prompting technique? I'm all…
I like the outputs from this model, but DAMN does it over think. Has anyone found a robust fix for this that isn't just capping output tokens? Anyone working on a 'thinking cap' for it? Some combo of llama params, or (system?) prompting technique? I'm all ears.
Collected discussion
I’m trying now by setting reasoning to low. Seems to address it, but need to test how it impacts, if at all, output quality. So far looks promising.
Llama.cpp is getting me the best speeds, with the most recent branch of tool calling fixes
You know what, I thought I saw in the docs that it was just off/high/max with high as default, I'll have to give Low a go
Don't second guess the trainer. Keep thinking high or max or off and call it a day
What are you serving on? VLLM will spiral into overthinking if you don't use the official deepseek v4 parsers. (For my setup claude originally tried to use deepseek R1)
i like how people call it reasoning lol
You've got choices: thinking off,low,high,max. Or you can limit the output. But that may result in truncated reasoning.