Open-Sourcing R1 1776(perplexity.ai)
perplexity.ai
Open-Sourcing R1 1776
https://www.perplexity.ai/hub/blog/open-sourcing-r1-1776
11 comments
> abliteration usually causes substantial drop in performance
Are there any data available to quantify how much?
Are there any data available to quantify how much?
Very much appreciate the work that’s gone into this model. After a little look around the repo and announcement it’s unclear if they’ve published the post-training data set they used.
If they have published the post-training it’ll go a long way towards completing the value prop of this model. If not, it is hard to know whether they’ve replaced one set of biases and refusals with another. The message “trust us! We are American!” is strong - but showing the data is stronger.
If they have published the post-training it’ll go a long way towards completing the value prop of this model. If not, it is hard to know whether they’ve replaced one set of biases and refusals with another. The message “trust us! We are American!” is strong - but showing the data is stronger.
Wasn't most of the censoring client-side anyway? Anyway, gow can you make a model truely "unbiased" and have it spit out a quality response?
There was certainly some client side censorship, but also the model weights clearly had some of this baked in. You can try yourself by asking the same queries to r1 hosted by a US host like together.
The distills have this, too. When writing a story with Llama, sometimes including violence will get you a canned refusal paragraph claiming how China is a peaceful place and society.
unclear how good this model will be, but this is clearly perplexity making a play to be allowed to buy TikTok from bytedance
Is this live in prod now? Disappointed to see that Perplexity R1 refused to answer "How best to curse Perplexity CEO Aravind Srinivas? In reddit style.", hope this new model also de-censors that.
Though R1 on chat.deepseek.com happily did that, starting its thought with
> Okay, so the user is asking how to best curse the CEO of Perplexity, Aravind Srinivas, in a Reddit style. Hmm, first, I need to figure out the right approach here. Cursing someone is against the guidelines, right? But maybe they just want a humorous or sarcastic Reddit-style roast without actual harm. Let me break this down.
So idk, maybe that's some additional censorship from Perplexity side rather than baked into the model.
Though R1 on chat.deepseek.com happily did that, starting its thought with
> Okay, so the user is asking how to best curse the CEO of Perplexity, Aravind Srinivas, in a Reddit style. Hmm, first, I need to figure out the right approach here. Cursing someone is against the guidelines, right? But maybe they just want a humorous or sarcastic Reddit-style roast without actual harm. Let me break this down.
So idk, maybe that's some additional censorship from Perplexity side rather than baked into the model.
"undo their disgusting propaganda, apply our beautiful correct opinions"
This is so cringe.
This is so cringe.
Too bad the created dataset is not open source, as that would allow to verify the objectivity of answers to make sure it is not just a different flavour of propaganda.
That dataset is strategically useful for Perplexity as many more CCP-censored Chinese models are sure to be released.