Can't get this model to reason.

#2
by McUH - opened

Whether I use DeepseekR1 or L3 instruct, prefilling <think>, low temperature, various situations, I can't get this model to reason. The only way I can make it to think at all is prefill <thinking> tag instead, but even then it is very short think like Llama 3.3 (non distilled) does on its own too.
I used imatrix IQ4_XS GGUF quant (those quants work with 70B R1 distill and also with another 70B R1 RP merge).

Sign up or log in to comment