What does it mean to reproduce DeepSeek R1-Zero? Like they have a model of equivalent performance? Is there a simple explanation of this post for those who aren't machine learning experts?
Also is the technique here related at all to the technique people think DeepSeek themselves used, where they apparently trained the model using OpenAI outputs?
R1-Zero is trained differently than most reasoning models, such as the "normal" R1 model, in regards what steps are done in training. TinyZero applies the same approach (but only on a subset of use cases) on a much smaller model to show it can apply on much smaller models as well.
The details of how it's trained different start to get into "machine learning expert" territory but you can get a decent high level via a casual read through of the DeepSeek link if you want to dive deeper.
I don't have a source to share, but I saw this claim on social media a few times in the last couple days, where people said their conversation with the model revealed that it thought it was some other OpenAI model. I have no idea how such training can work using another model's output, but I saw comments claiming that this is why their training was so cheap.
Also is the technique here related at all to the technique people think DeepSeek themselves used, where they apparently trained the model using OpenAI outputs?