Enable Pipeline arguments for the models on the widget

We have uploaded (still private) a model fine-tuned on a high quality Named Entity Recognition (NER) private dataset. Although the dataset is better than conll2002, in the examples like the one from here the widget provides weird values like “…Wolf (PER) gang (O)…”. This should be fixed with the pipeline parameters, more concretely with the aggregation_strategy parameter.

I thoroughly checked this website and I have not found any way to alter this default pipeline parameter value. Maybe you could tackle this by introducing a new parameter to the widget.

By the way, I do not see the widget parameter here: https://github.com/huggingface/huggingface_hub/blame/main/modelcard.md

[julien-c]

i think we can definitely support this use case, and I think the way to go here would be to make the API calls configurable through the inference: attribute in the model card metadata.

So something like:

inference:
  parameters:
    aggregation_strategy: "none"

(widget: sounds not appropriate to me because it controls the frontend, not the pipeline/API, and besides it would be hard to extend the current data type (an array of strings or objects to be used as simple inputs). inference: is already defined but can only be false currently (to turn off the inference + widget))

The parameters would then be set at the model repo level and not necessitate any frontend change (apart from tweaking the API call)

Thoughts?

(we should make sure that those options are documented consistently btw)

[mishig25]

@julien-c yes I’ll work on this 👍