Open source is a very loose term when it comes to GenAI. Like Llama the weights are available with few restrictions but importantly how it was trained is still secret. Not being reproducible doesn’t seem very open to me.
Their paper outlines the training process but doesn’t supply the actual data or training code. There is a project on huggingface: https://huggingface.co/blog/open-r1 that is attempting a fully open recreation based on what is public.
Open source means it can be publicly audited to help soothe suspicion, right? I imagine that would take time, though, if it’s incredibly complex
Open source is a very loose term when it comes to GenAI. Like Llama the weights are available with few restrictions but importantly how it was trained is still secret. Not being reproducible doesn’t seem very open to me.
True, but in this case I believe the also open sourced the training data and the training process.
Their paper outlines the training process but doesn’t supply the actual data or training code. There is a project on huggingface: https://huggingface.co/blog/open-r1 that is attempting a fully open recreation based on what is public.