r1-1776 Discussions
R1 1776 is a DeepSeek-R1 reasoning model that has been post-trained by Perplexity AI to remove Chinese Communist Party censorship. The model provides unbiased, accurate, and factual information while maintaining high reasoning capabilities.
Or in other words R1 1776 is DeepSeek-R1 finetuned on American instead of Chineese propaganda which is really upsetting the Chineese AI community. The model is currently getting flooded with malicious reports and already received 13 reports. I highly doubt that HuggingFace as a primary American company will take it down but just to be sure I downloaded it to my own subvolume on gpool and so can convert it to GGUF at any time you like.
URL: https://huggingface.co/perplexity-ai/r1-1776
There are two main issues with quantizing this model:
- https://github.com/ggml-org/llama.cpp/pull/11446 is not merged yet so we would need to requant it once it is which is really painful for such a massive model
- With CastlePeak my RPC setup is currently partialy down and I will not receive the new PSU this week. Maybe I could temporarily move on of the PSUs from StormPeak to CastlePeak as StormPeak has 2x 1500 Watt PSUs but even then I managed to let a piece of plastic fall into radiator of the AIO watercooling system of CastlePeak so there are a lot of uncertainties to when I could get an RPC setup with more than 512 + 128 GB of RAM operational.
I just ordered a new AiO cooling solution for CastlePeak which will arrive tomorrow evening. So the only remaining issue to get the full RPC setup working is finding a PSU.
I found an old 600 Watt PSU. It technically is not powerful enough As TDP CPU + TDP GPU is already 600W but if I set a lower CPU clock speed and limit the GPU power it will probably work especially given how we barely use CPU and GPU during imatrix computation. If everything goes well we could have a full imatrix setup by late evening tomorrow.
Well done ppl-ai. And you :)
Until somebody gives me numbers that prove otherwise, I have zero qualms using Q4_K_M or even lower to make imatrices :)
And ... I see your efforts with great joy but also big trepidation :) The queue situation is really bad, and gets worse. The only reason it doesn't look like disaster is because I started mitigating by not imatrix-quanting models I would normally imatrix.
But realistically, since we woulds have to redo it anyway, and I am sure it will get merged, I don't think it makes sense to quant it now, at least not with a full set of quants.
Or in other words R1 1776 is DeepSeek-R1 finetuned on American instead of Chineese propaganda
Yea, this is something that also came to my mind when I heard about it. I somehow have doubts about it really being "uncensored". My expectations towards this model are quire low as of now. I suspect it's just not-politically-biased and that its "uncensored" aspect won't really affect matters that common people may actually want to be uncensored, unbiased and raw. I'm expecting it to trip over guidelines in first response lol.