Perplexity AI has introduced R1 1776, a groundbreaking iteration of the DeepSeek-R1 reasoning model, specifically post-trained to eliminate Chinese Communist Party censorship influences. This enhancement ensures the model delivers unbiased, accurate, and factually correct information without compromising its sophisticated reasoning abilities. To validate the model’s uncensored and comprehensive engagement with sensitive topics, a diverse and multilingual evaluation set comprising over 1000 examples was employed. This set, alongside human annotators and LLM judges, was crucial in assessing the model’s ability to handle and respond to a wide array of queries without resorting to evasion or overly sanitized responses. Moreover, rigorous evaluations confirmed that R1 1776 retains the original model’s mathematical and reasoning prowess, indicating that the decensoring process did not detract from its core functionalities.
Read more…
