metadata_update() function inserts verified=True in self-reported evaluations

[lewtun]

Describe the bug

With the refactor of evaluation metadata in #940, it seems that metadata_update() can no longer differentiate between self-reported evaluations and those from Hugging Face's evaluation service.

As a result, updating a model which has both types of evaluation metrics will result in a verified=True entry getting inserted into the self-reported metrics. This is undesirable, because the verified field is used to distinguish both types of evaluations on leaderboards etc.

Another effect from #940 is that evaluations with the same task type / name are merged together, which isn't ideal because logically a self-reported evaluation and an automated one should each have their own entry in the model-index.results array.

I'll take a stab at fixing this if @Wauplin doesn't beat me to it :)

cc @abhishekkrthakur

Reproduction

Run the following snippet to open a Hub PR and see the insertion of the verified=True field:

from huggingface_hub import metadata_update

model_id = "autoevaluate/binary-classification"
card = ModelCard.load("autoevaluate/binary-classification")
metadata = card.data.to_dict()
metadata_update(model_id, metadata=metadata, overwrite=True, create_pr=True, commit_message="Test update")

Logs

No response

System info

- huggingface_hub version: 0.11.0.rc0
- Platform: Linux-4.19.0-22-cloud-amd64-x86_64-with-glibc2.10
- Python version: 3.8.13
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Token path ?: /home/lewis_huggingface_co/.huggingface/token
- Has saved token ?: True
- Who am I ?: lewtun
- Configured git credential helpers: store
- FastAI: N/A
- Tensorflow: N/A
- Torch: 1.11.0
- Jinja2: 3.1.2
- Graphviz: N/A
- Pydot: N/A