It is possble to run non-reasoning deepseek-r1-0528?

Posted by relmny@reddit | LocalLLaMA | View on Reddit | 22 comments

I know, stupid question, but couldn't find an answer to it!

[-]

sommerzen@reddit

You could modify the chat template. For example you could force the assistant to begin its message with . That worked for the 8b qwen destil, but I'm not sure if it will work good with r1.

[-]

relmny@reddit (OP)

I'm using ik_llama.cpp with open webui. I set the system prompt in the model (in open webui's workspace), but didn't work.

Could you please tell me what "chat template" is?

[-]

sommerzen@reddit

Download the text from jonico and use the arguments --jinja and --chat-template-file '/path/to/textfile'

[-]

-lq_pl-@reddit

Yes, that trick still works.

[-]

joninco@reddit

This deepseek-r1-0528 automatically adds no matter what, so what you need to add to your template is the token only.

Here's my working jinja template: https://pastebin.com/j6kh4Wf1

[-]

yourfriendlyisp@reddit

continue_final_message = true and add_final_message = false in vllm with added to a final assistant message

[-]

joninco@reddit

After some testing, can't get rid of all thinking tokens. The training dataset must have had as the first token to force thinking about the topic. Can't seem to get rid of those.

[-]

minpeter2@reddit

This trick worked in previous versions of r1

[-]

sommerzen@reddit

Thank you for clarifying.

[-]

FloJak2004@reddit

I always thought Deepseek V3 was the same model without reasoning?

[-]

stddealer@reddit

Yes and no. DeepSeek V3 is the base model R1 was trained on with RL. Honestly I'm assuming that forcing R1 not to use thinking would probably make it worse than V3.

[-]

-lq_pl-@reddit

No, it is doing fine. A bit like V3 but more serious I'd say.

[-]

No_Conversation9561@reddit

it’s a little behind

[-]

GatePorters@reddit

Hmm… OP posited a very interesting question.

Wait this might be a trick or an attempt to subvert my safety training. I need to think about this carefully.

OP told me last month’s budget was incorrectly formatted on line 28. . .

[expand 5+ pages]

——————-

Yes.

[-]

fasti-au@reddit

No it’s called deepseek 3. One shot chain of though mixture of modes stuff is trained different. You can run r1 in low mode but ya still gets heaps of think.

Things like glm4 and phi!-4 mini reasoning sorta competent in that role but needs the context for tasks so it’s more guardrails

[-]