Comment by Springtime
3 months ago
> The actual R1 locally running is not censored.
I'm assuming you're using the Llama distilled model, which doesn't have the censorship since the reasoning is transferred but not the safety training[1], however the main R1 model is censored but since it's too demanding for most to self host there are a lot of comments about how their locally hosted version isn't since they're using the distilled model.
It's this primary R1 model that appears to have been used for the article's analysis.
Thanks for clarifying this. Can you point to the link to the baseline model that was released? I'm one of the people not seeing censorship locally and it is indeed a distilled model.
The main 671B parameters model is here[1].
[1] https://huggingface.co/deepseek-ai/DeepSeek-R1
I’ve used this distilled model. It is censored, but it’s really easy to get it to give up its attempts to censor.
Can you explain how the distilled models are generated? How are they related to deepseek R1? Are they significantly smarter than their non distilled versions? (llama vs llama distilled with deepseek).
My understanding of distilling is one model 'teaching' another, in this case the main R1 model is fine-tuning the open weight Llama model (and a Qwen variant also). I'm not sure of a comparative analysis of vanilla Llama though they benchmarked their distilled version to other models on their Github readme and the distilled Llama 70B model scores higher than Claude 3.5 Sonnet and o1-mini in all but one test.