Perplexity AI has open-sourced R1 1776, a model of the DeepSeek-R1 language mannequin that has been post-trained to get rid of censorship and supply factual responses. Whereas the mannequin weights can be found on Hugging Face, the mannequin will be accessed through Sonar API.
“We’re contemplating open-sourcing coaching and inference code as effectively. Not determined but, however neighborhood and open supply is one thing we intend to do extra of since our worth goes to be in offering an awesome assistant and personalising it to the consumer, not within the fashions themselves,” Perplexity AI chief Aravind Srinivas mentioned.
DeepSeek-R1 is an open-weight massive language mannequin (LLM) with reasoning capabilities just like these of main fashions like o1 and o3-mini.
Nevertheless, its unique model has been famous for refusing to interact with sure delicate matters, significantly these censored by the Chinese language Communist Celebration (CCP). Perplexity’s post-training effort focuses on mitigating this concern.
One such instance highlighted by Perplexity entails a question about Taiwan’s independence and its potential impression on NVIDIA’s inventory. DeepSeek-R1 initially responded with CCP-aligned statements, avoiding any direct evaluation. In distinction, R1 1776 now gives an in depth response, outlining the geopolitical and financial dangers that would have an effect on NVIDIA’s inventory worth. It discusses potential provide chain disruptions, market volatility, geopolitical retaliation, navy battle dangers, and regulatory shifts.
Perplexity’s post-training course of included gathering a dataset of 40,000 multilingual prompts centered on censored matters.
“We employed human specialists to establish roughly 300 censored matters,” Perplexity AI mentioned in a weblog submit. A multilingual censorship classifier was developed to filter queries, guaranteeing that responses have been factual in addition to related. The crew used NVIDIA’s NeMo 2.0 framework to refine the mannequin whereas sustaining its reasoning capabilities.
To guage the effectiveness of R1 1776, Perplexity examined it on a dataset of over 1,000 examples overlaying a broad vary of delicate matters. The corporate employed each human annotators and LLM judges to evaluate whether or not the mannequin would evade responses or present overly sanitised solutions.
“Our evaluations present that the mannequin stays absolutely uncensored and performs on par with the bottom R1 mannequin in reasoning and mathematical benchmarks,” Perplexity reported.
Perplexity AI not too long ago introduced that its in-house mannequin, Sonar, might be obtainable to all Professional customers on the platform. Customers with the Perplexity Professional plan could make Sonar the default mannequin through settings.
The corporate additionally launched Deep Analysis, a software for autonomously conducting in-depth analysis and evaluation. The characteristic performs a number of searches, critiques lots of of sources, and compiles findings into complete experiences. It’s free for everybody, as much as 5 queries per day for non-subscribers and 500 queries per day for Professional customers.
The submit Perplexity AI Open-Sources R1 1776 to Take away Censorship from DeepSeek-R1 appeared first on Analytics India Journal.
