mradermacher

/

model_requests

like 71

Let's avoid lmstudio-community. The original models are:

• https://huggingface.co./deepseek-ai/DeepSeek-R1-Distill-Llama-8B
• https://huggingface.co./deepseek-ai/DeepSeek-R1-Distill-Llama-70B

Not compleately sure if llama.cpp supported but I seem some GGUFs of them already exist so they might be.

requesting an uncensored version please.

We just quantizing and not finetuning. I personally kind of specialized myself into creating uncensored versions of models so I will likely try to create one if the model is compatible with axolotl. This is a manual and somewhat time consuming process so let’s hope I file fine time to do so.

Let's avoid lmstudio-community. The original models are:

• https://huggingface.co./deepseek-ai/DeepSeek-R1-Distill-Llama-8B
• https://huggingface.co./deepseek-ai/DeepSeek-R1-Distill-Llama-70B

Not compleately sure if llama.cpp supported but I seem some GGUFs of them already exist so they might be.

requesting an uncensored version please.

We just quantizing and not finetuning. I personally kind of specialized myself into creating uncensored versions of models so I will likely try to create one if the model is compatible with axolotl. This is a manual and somewhat time consuming process so let’s hope I file fine time to do so.

Thanks,
I've seen that models with qwen architecture give better results, would u consider uncensoring the 7B qwen version?
I'm sure someone will eventually consider doing it, but thanks for considering it

Should all be queued by now

I've seen that models with qwen architecture give better results, would u consider uncensoring the 7B qwen version?

@Ein69420 I started uncensoring DeepSeek-R1-Distill-Qwen-7B as requested. It should be done in around 3 hours. Let's hope the model turns out well.

so, repo when? :-)

so, repo when? :-)

I just compleated it.
@mradermacher Please quant it: https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored

@Ein69420 GGUF quants of DeepSeek-R1-Distill-Qwen-7B-Uncensored should be avilable soon. If you want to use the Lora in the meantime I uploaded it to https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored-Lora

well done :)

I will further finetune DeepSeek-R1-Distill-Qwen-7B-Uncensored on https://huggingface.co./datasets/Guilherme34/Reasoner-dataset-FULL-roles and upload it as a separate model so we have an uncensored model with similar or better reasoning capabilities than the original.

DeepSeek-R1-Distill-xxx works well (some sort) only in English and Chinese. Tried other languages, it wildly hallucinated, sounds likes it completely forgotten other languages which were originally pretrained/finetuned by Meta or Qwen.

Steve
EraX AI.

so, repo when? :-)

I just compleated it.
@mradermacher Please quant it: https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored

@Ein69420 GGUF quants of DeepSeek-R1-Distill-Qwen-7B-Uncensored should be avilable soon. If you want to use the Lora in the meantime I uploaded it to https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored-Lora

NICE!
danke!

@Ein69420 I now created https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored-Reasoner which should be the best uncensored DeepSeek R1 Distill finetune as it not only uncensored it but after it adds back its reasoning capabilities. That one took 2x RTX 4090 for around 10 hours to train. If you don't want to use step-by-step reasoning, I recommend sticking with the uncensored one but that would be against the spirit of the DeepSeek R1 Distill series of models designed for reasoning. A quantized version should be released by mradermacher soon but imatrix quants might get delayed due to DeepSeek-R1 RPC imatrix computation. If you want to use the Lora instead you can find it under https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored-Reasoner-Lora

@mradermacher Please quantize https://huggingface.co./nicoboss/DeepSeek-R1-Distill-Qwen-7B-Uncensored-Reasoner

It's queued