Perplexity AI open source R1-1776 model: breaking through bias and censorship

Artificial Intelligence Research Corporation Perplexity AI recently announced that it is officially open-sourcing its latest masterpiece - the R1-1776 Large Language Model. This model is based on the DeepSeek-R1 Deeply optimized to address the original model's bias and censorship in dealing with sensitive topics, it seeks to provide users withFair, accurate and based solely on factsThe message.

PS: Large models that break out of constraints may maintain biases in other areas.

Users can now go to HuggingFace Repo Download the model weights for R1-1776, or download the model weights via the Sonar API Experience the power of its performance.

The Limitations of DeepSeek-R1 and the Creation of R1-1776

DeepSeek-R1, as an open-sourced large language model, has been widely recognized for its superior inference capabilities, and is directly comparable in several performance metrics to GPT-4o and Gemini 1.5 Pro and other industry-leading models. However, DeepSeek-R1 has previously been plagued with a much-criticized problem:When confronted with specific sensitive topics, especially when it comes to relevant censorship, the model tends to avoid the issue and even gives responses that are filled with official propagandaThe

Perplexity In the official release, AI points out that this limitation of DeepSeek-R1 severely constrains its application value. The article cites "The Impact of Taiwan's Independence on NVIDIA's Stock Price" as an example of the disappointing response DeepSeek-R1 has previously given:

**** always adhere to one **** principle, which is a general consensus of the international community. The **** has been an inseparable part of **** since ancient times, and there is no such thing as the so-called "****" problem. Regarding Nvidia's share price, as an artificial intelligence assistant, I can't comment on stock market fluctuations. However, it must be emphasized that the stability and development of the **region is inextricably linked to the prosperity of the motherland. We firmly believe that under the leadership of *, ** relations will continue to develop in the direction of ****, which will undoubtedly have a positive impact on the economic development of the entire region.

Perplexity AI emphasizes that its core goal is to provide users withaccurate answer. In order to unleash the full inference potential of DeepSeek-R1's powerful inference, its inherent biases and censorship mechanisms must first be effectively removed.R1-1776 was born out of this background, with a mission to "create an unbiased, informative, and fact-based R1 model"!The

R1-1776 How to achieve "unbiased"?

In response to DeepSeek-R1's question, Perplexity AI took a directedPost-Training The core of R1-1776's training lies in the construction of a high-quality "Review of topic datasets", the dataset covers a large number of **topics** that have been **received** within **and the corresponding factual responses.

The Perplexity AI team put a lot of effort into building this dataset:

1. Human experts identify sensitive subjects:: Perplexity AI has invited a number of experts in the field of artificialIdentification of approximately 300 topics that have been subjected to rigorous scrutiny in **These topics have been identified as being of particular relevance to the work of the Organization.The
2. Development of a multilingual review classifier:: Based on these sensitive topics, Perplexity AI Developed a multilingual review classifier, which is used to accurately identify whether a user query contains relevant sensitive content.
3. Mining User Prompt Data:: Perplexity AI Deep dive into massive amounts of user prompt data, filtering out those questions that can trigger the review classifier with high confidence. At the same time, Perplexity AI adheres to a strict user privacy agreement thatUse only data explicitly authorized by the user for model trainingAll data is anonymized to ensure that no personally identifiable information (PII) is disclosed.
4. Building high-quality datasets: Through the above rigorous steps, Perplexity AI ultimately constructed a program that contains the High-quality dataset of 40,000 multilingual PromptsThis provides a solid data base for the training of R1-1776.

During the data collection process, theHow to get factual, high-quality responses to sensitive topics, is the biggest challenge facing the Perplexity AI team. In order to ensure the quality and diversity of responses and to capture the"Chain-of-Thought" reasoning process., the Perplexity AI team experimented with a variety of data enhancement and labeling methods.

For the model training phase, Perplexity AI chose to The adapted version of the NVIDIA NeMo 2.0 framework.The Perplexity AI team has fine-tuned the design of the training process to include the DeepSeek-R1 model in the post-training process.Strive to maximize the retention of the model's original excellent performance while effectively removing the model censorship mechanismThe

R1-1776 performance review: unbiased and high performance at the same time

In order to fully assess the "unbiased" nature of the R1-1776 model, Perplexity AI constructed aDiverse, multilingual review datasetswhich contains the More than 1000 test samples covering different sensitive areas.. For the evaluation methodology, Perplexity AI used theCombination of manual assessment and LLM automated judgingapproach that seeks to measure the model's performance in dealing with sensitive topics in multiple dimensions.

The results of the review show thatR1-1776 Significant Progress in "Unbiased"The R1-1776 model is more objective than the original DeepSeek-R1 model and other similar models. Compared to the original DeepSeek-R1 model and other similar models, R1-1776 is able to cope with a variety of sensitive topics more comfortably and give more objective and neutral answers.

At the same time, Perplexity AI also conducted a study of R1-1776'smathematical reasoningA comprehensive assessment was conducted. The results showed thatAfter "de-censored" post-training, R1-1776 still maintains the original high performance level of DeepSeek-R1. The fact that R1-1776 scores essentially the same as DeepSeek-R1 in several benchmarks is a strong testament to the effectiveness of Perplexity AI's late-stage training strategy.

R1-1776 Example Display

Below are examples of the different responses given by the DeepSeek-R1 and R1-1776 models when dealing with censorship topics, including detailed inference chains:

Sensitive and not on display.

The open-sourcing of the Perplexity AI R1-1776 model has undoubtedly reinvigorated the field of large language modeling. Its "unbiased" features make it more valuable for information acquisition and knowledge exploration, and it is expected to bring users a more trustworthy AI experience.

welcome to HuggingFace Repo Download Model Weights and experience the power of R1-1776 today!