Perplexity Open-Sources R1 1776 Model for Unbiased Information

Perplexity has open-sourced R1 1776, a version of the DeepSeek-R1 model, to provide uncensored and factual information. The model is available for download on HuggingFace and can also be accessed via the Sonar API according to a company blog post.

DeepSeek-R1, known for its reasoning capabilities, faced limitations due to its refusal to address sensitive topics, particularly those censored by the Chinese Communist Party. To overcome this, Perplexity post-trained the model to ensure it delivers unbiased and accurate responses. This involved gathering a dataset of 40,000 multilingual prompts and employing human experts to identify censored topics detailed in a company blog post.

The post-training process utilized Nvidia's NeMo 2.0 framework to maintain the model's reasoning abilities while removing biases. Evaluations showed that R1 1776 performs on par with the original R1 model, ensuring its effectiveness in handling a wide range of sensitive topics.