It’s been scarcely a week since DeepSeek’s launch of their R1 AI reasoning model, and we have already seen researchers at Hugging Face aspiring for a more transparent version of this technology. Their pursuit, labelled “Open-R1,” is all about ‘open knowledge.’
Hugging Face’s research head, Leandro von Werra, along with a team of engineers, is venturing on this mission to recreate an open-source version of the R1 model, including the dataset utilized for its training. DeepSeek’s secrecy around the tools used to develop the R1 model have led many industry professionals to call their approach a “black box” release philosophy.
Noted by Hugging Face engineer, Elie Bakouch, open-sourcing the R1 architecture allows more than just clarity to users, it also unlocks the potential for further research and innovation.
Replicating R1 model is not as smooth as it seems. The Chinese AI lab DeepSeek, that produced it, has managed to create a model that surpasses the performance of the OpenAI’s o1 reasoning model on many benchmarks. Yet, without the transparency of training or instructions provided, it’s arduous for researchers to study and eventually recreate it.
The Open-R1 project is relying heavily on their Science Cluster, a dedicated research server to generate datasets similar to those used by DeepSeek for training. The expectation is not just to clone the R1 model but also lay a sturdy foundation for the upcoming models.
While open-source AI comes with its set of potential misuse worries, the benefits it provides when it comes to knowledge dispersal seem to outweigh its downsides considerably. Hence, fostering the idea that it’s not just a handful of labs capable of progress, but open source could also lead the way.
Original source: Read the full article on TechCrunch
