Add datasets to model card

[ittailup]

In the spirit of #1004 and #1005, this automatically adds a "datasets" object to the model card through the "dataset_tags" attribute.

The reason we use dataset_tags and not dataset comes from here

Datasets are from model card if they are local directories, using the same method as in https://github.com/OpenAccess-AI-Collective/axolotl/blob/dec66d7c53a2de6cf74911faf9c1ad1f7f0fff14/src/axolotl/utils/data.py#L244

Running @hamelsmu tiny-mistral config from here we get the following model card:

---
library_name: peft
tags:
- generated_from_trainer
datasets:
- mhenrichsen/alpaca_2k_test
base_model: openaccess-ai-collective/tiny-mistral
model-index:
- name: temp_dir
  results: []
---

[winglian]

I like this, I do think we should try to figure out if a path is a local path or hf dataset though.

[ittailup]

checking if local dir now, should be good. I tried to also check if we were saving the tokenized dataset to hfhub too, but when I enabled push_dataset_to_hub I would get

huggingface_hub.utils._validators.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': 'ittailup/alpaca-test/942a2dcbc346823803788fd18b4bb1bf'. Use `repo_type` argument if needed.

will report as separate issue.

[winglian]

I'm happy to fix the linti g tomorrow for you.

[hamelsmu]

@winglian I fixed the lint (my fault) in #1014

[hamelsmu]

Actually I'll fix the lint here too, its a small commit nevermind I can't push to this PR

[osanseviero]

Very cool!

[osanseviero]

Very cool! Given you're adding datasets, here you can potentially also add finetuned_from. This is equivalent to base_model in the configuration from axolotl and will link the fine-tuned models to their base models

See https://huggingface.co/docs/hub/model-cards#specifying-a-base-model

[osanseviero]

Actually this might already be specified 🔥

[ittailup]

Yes! I added it in my first try but once I saw the output and code I noticed it was already getting picked up (you can see it in hamel's example of the model card raw print.

I am wondering what other things could get pushed on to the model card metadata as a separate function. Perhaps tags could get auto-added from specific settings? I've created the obvious one (evals) at #1020 . The other one mentioned on the issue is CO2 usage, but could be a little bit more difficult.

[winglian]

@osanseviero should we use something like hf_api.list_models(filter=...) to verify that the base_model specified isn't a locally downloaded model? or is there a simpler lightweight check?

[osanseviero]

You can go with huggingface/huggingface_hub#36 (comment) (so catching an error upon doing model_info) - that would likely be the most precise

[winglian]

🙌

[winglian]

So I just realized this only handles the case where it creates a model card when no hub_model_id is defined and it doesn't get pushed to hf. We need to add the dataset tags here as well for the case when the model gets automatically pushed to the hub and the model card gets created in that step.

[winglian]

@ittailup do you want me to help you with making the changes?