A research team from Hugging Face has announced the Open-R1 project, aiming to train a new model equivalent to DeepSeek-R1 but with a dataset open for replication by others.
The core principle of DeepSeek-R1 lies in the straightforward approach of providing reasoning data beforehand to tackle problems. Subsequently, they attempt to train the LLM model normally, rewarding reasoning via reinforcement learning (RL) similar to rewarding AI in gaming.
Although DeepSeek-R1 opens the model running code to the public, they do not disclose the dataset, nor the training code, limiting the model’s study for practical use only. Creating an open dataset alongside training code accessible to all would significantly broaden the model’s developmental scope.
The development direction of Open-R1 is divided into 3 steps:
1. Create a reasoning data set using output data from DeepSeek-R1 itself.
2. Develop RL training code capable of training other LLM models to provide similar reasoning.
3. Demonstrate creating a model similar to DeepSeek-R1 in full form.
If the team successfully creates the dataset and demonstrates model training, it will pave the way for developing other reasoning models, such as medical reasoning where models can think through diagnoses before predicting diseases or recommending treatments.
As of now, the project is still in its early stages but has garnered thousands of stars on GitHub.
TLDR: Hugging Face’s research team introduces the Open-R1 project, aiming to train a new model akin to DeepSeek-R1 with an open dataset for others to replicate, potentially expanding the scope for developing reasoning models, including in medical diagnostics.
Leave a Comment