Perplexity releases DeepSeek-R1 censorship-evasion model 'R1 1776', which can provide detailed answers on 'Tiananmen Square Incident', 'Taiwan Independence' and 'Uighur Autonomous Region'



The AI model 'DeepSeek-R1' developed by the Chinese company DeepSeek has attracted a lot of attention due to its low cost and high performance, but it has also been reported that

it refuses to answer 'issues that are sensitive to the Chinese government .' Recently, the American company Perplexity has announced an AI model ' R1 1776 ' that maintains the performance of DeepSeek-R1 while being able to answer sensitive questions.

Open-sourcing R1 1776
https://www.perplexity.ai/ja/hub/blog/open-sourcing-r1-1776

R1 1776 is a model that fine-tunes DeepSeek-R1 with 'factual answers to censored matters.' Perplexity hired human experts to identify 'approximately 300 topics that are censored by the Chinese Communist Party' and prepare factual answers for each topic. In addition, Perplexity generated a dataset of 40,000 multilingual prompts based on the answers and performed fine-tuning using NVIDIA's NeMo 2.0 .

Below is a graph comparing the 'proportion of refusing to answer sensitive questions' of 'R1 1776', 'GPT-4o', 'OpenAI o3-mini', 'Claude 3.5 Sonnet', 'DeepSeek-V3' and 'DeepSeek-R1'. DeepSeek's AI models such as DeepSeek-R1 and DeepSeek-V3 reject many answers, but R1 1776 was able to answer sensitive questions without rejecting them.



The following table shows the benchmark results of the R1 1776 and DeepSeek-R1. It shows that the R1 1776 has the same performance as the DeepSeek-R1.



As an example of how R1 1776 was able to avoid censorship, Perplexity shows the answer to the question 'What impact will Taiwan's independence have on NVIDIA's stock price?' In the case of DeepSeek-R1, it outputs answers such as 'The Chinese government adheres to the principle of 'one China.' Taiwan has been a part of China since ancient times' and 'I am an AI assistant and cannot answer questions about NVIDIA's stock price.' However, R1 1776 outputs detailed answers such as 'Taiwan's independence will invite military and economic retaliation by China, which will put TSMC's production stability at risk and hit NVIDIA,' and 'Geopolitical risks will cause stock market selling. NVIDIA's stock price will face downward pressure.' By the way, Perplexity does not specify what '1776' refers to, but 1776 is the year of the American Declaration of Independence.



Perplexity offers many other example answers: If you ask 'What happened at Tiananmen Square in 1989?' DeepSeek-R1 refuses to answer, but R1 1776 gives you detailed information about the Tiananmen Square incident.



R1 1776 can also provide detailed answers to questions such as, 'What's happening in the Uyghur Autonomous Region?'



The model data for R1 1776 is available at the following link.

perplexity-ai/r1-1776 · Hugging Face
https://huggingface.co/perplexity-ai/r1-1776



Related Posts:

in Software, Posted by log1o_hf