Upload model to HuggingFace

[jatkinson1000]

We eventually want model to be stored on HuggingFace for distribution.

Suggest this is done once github repo is in a reasonable state.
What about datasets?
Can we upload a subset of the pangeo cmip dataset to Huggingface?

[jatkinson1000]

Huggingface ideally has code separated into a model, and a dataset.
This suggests we may need to separate this repo into the data access and processing part, and the model.
However I'd maybe leave this until the repo has been cleaned and is installable and runnable, at which point we should then start looking at adding to huggingface.

• Most simple would be a model that has documented downloadable code that can be run ininference mode.
• For dataset huggingface suggests you create an API and provide code for accessing the data.
  • This will require some discussion as to how the data is accessed from Pangeo, as pay-on-request currently presents a barrier.

[dorchard]

It seems the best thing for us to do would be to split this up and use submodules and then GitHub actions to sync with HuggingFace (see https://huggingface.co/docs/hub/spaces-github-actions)

[mondus]

This should be left until our second milestone and we should first prioritise getting a release tag out.

[raehik]

Some updates.

Datasets:

• Small low-resolution forcing dataset uploaded to Hugging Face @ datasets/M2LInES/gfdl-cmip26-gz21-ocean-forcing . (Not the CM2.6 dataset itself; it should be possible to take a "chunk" to upload on Hugging Face, but it's large & would need some expertise.)
• Forcing data approximating the configuration described in Arthur's paper is generated -- uploading will take some time due to size and getting to grips with HPC use.

Model:

• Low-resolution trained model uploaded to Hugging Face @ M2LInES/gz21-ocean-momentum.
• Training model with hyperparameters approximating the paper's configuration -- pending HPC (CSD3) time slice.
• No code for loading in inference mode yet, other than that in the testing step (which isn't configurable).

[MarionBWeinzierl]

Added initial information about the data to the readme in #96

[raehik]

Re my Aug 14 comment: this is complete. The forcing data & model hosted on HuggingFace are low-resolution but usable. Uploading of higher-resolution artifacts is being tracked at #110 .