Uncensored DeepSeek-R1 : Perplexity-ai R1–1776
DeepSeek-R1 without Chinese bias and censorship
DeepSeek blew the Generative AI market towards late January this year, even breaking the US Stock market. Though the model is great, it has major flaws
Problem with DeepSeek-R1
https://medium.com/media/da9f764b7dde5f07cffcbef4f6ee9973/href
- Bias and Censorship: The original DeepSeek-R1 model refuses to address sensitive topics, particularly those censored by the Chinese Communist Party (CCP). For instance:
When asked about Taiwan’s independence impacting Nvidia’s stock price, it delivers CCP-aligned responses (e.g., “Taiwan has been an integral part of China since ancient times”) and avoids substantive analysis.
- Limitation: This censorship limits its utility for providing accurate answers to all user queries, a core goal of Perplexity.
To counter this bias, Perplexity-AI has now launched a new fine-tuned DeepSeek-R1, R1–1776, free from any bias and censorship.
https://medium.com/media/50bb071340a0655953bdb13503b545a8/href
Overview of R1–1776
- Purpose: R1–1776 is a post-trained version of the DeepSeek-R1 model, designed to provide unbiased, accurate, and factual information.
- Availability: Open-sourced by Perplexity, with model weights downloadable from their HuggingFace repository or accessible via the Sonar API.
- Origin: Built upon DeepSeek-R1, a fully open-weight large language model (LLM) with reasoning capabilities close to state-of-the-art models like o1 and o3-mini.
Improvements in R1–1776
- Objective: Mitigate bias and censorship while preserving DeepSeek-R1’s powerful reasoning capabilities.
- Post-Training Focus: Address censored topics, especially those restricted by the CCP, to ensure factual and unbiased responses.
Post-Training Process
- Data Collection:
Identified ~300 CCP-censored topics using human experts.
Built a multilingual censorship classifier.
Mined 40,000 user prompts triggering censorship, ensuring user consent and no PII.
- Response Generation: Gathered factual responses with valid chain-of-thought reasoning, overcoming challenges in sourcing high-quality completions.
- Training Framework: Used an adapted Nvidia NeMo 2.0 framework to “de-censor” the model while maintaining performance on academic and internal benchmarks.
Metrics
As you can observe, the new R1–1776 has almost 0 Chinese bias compared to 80 for DeepSeek-R1
Even on baseline evaluation metrics, the model is neck-neck with the original DeepSeek-R1.
Example outputs
How to use Perplexity-AI R1–1776?
The model is open-sourced and the weights can be explored at HuggingFace
Hope you try out the uncensored version of DeepSeek-R1
Uncensored DeepSeek-R1 : Perplexity-ai R1–1776 was originally published in Data Science in your pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
