thinking-machines-lab
Consider adding `push_to_hub()` functionality for trained models
Hello Thinking Machines team, congrats on the release of Tinker and the wonderful docs and clean API!
I was wondering if you'd be open to adding a push_to_hub() functionality to Tinker, where users can choose to push their model checkpoints to the Hugging Face Hub (either mid-training or afterwards)?
If your models are implemented as transformers models, this should be pretty straightforward and you can check out e.g. how we do this in TRL here to also supply a pretty model card, complete with metadata about the dataset lineage, how to run the model etc.
One nice aspect of this is that you can add e.g. a tinker tag and banner and get visibility on the Hugging Face Hub for models trained with Tinker. (TRL example below)
If this is a feature that you're interested in supporting, please let me know if you need any pointers!
There are also some additional docs on adding a model Hub integration in https://huggingface.co/docs/hub/models-adding-libraries#integrate-your-library-with-the-hub, which might be helpful to.
