Today we're open-sourcing R1 1776—a version of the DeepSeek R1 model that has been post-trained to provide uncensored, unbiased, and factual information.
To keep our model "uncensored" on sensitive topics, we created a diverse, multilingual evaluation set of 1000+ examples. Using human annotators and specially designed LLM judges, we compared frequency of censorship in the original R1 and state-of-the-art LLMs to R1 1776.
We also ensured that the model’s math and reasoning abilities remained intact after the uncensoring process. Benchmark evaluations showed it performed on par with the base R1 model, indicating that uncensoring had no impact on core reasoning capabilities.
DeepSeek-R1 rivals top reasoning models like o1 and o3-mini. However, its usefulness is limited by its refusal to engage with topics censored by the CCP. We aim to always provide accurate answers, but had to address R1's censorship before using its reasoning capabilities.
