Hi,
We have recently released a Norwegian mBERT-base based model finetuned on MNLI but noticed the results of the widget on the model card page were not very accurate. After some digging, it seemed it was using the default English hypothesis template ("This example is {}."). When we tried changing its value in the model card to a proper Norwegian hypothesis template it just did not work.
pipeline_tag: zero-shot-classification
widget:
- text: "Folkehelseinstituttets mest optimistiske anslag er at alle voksne er ferdigvaksinert innen midten av september."
candidate_labels: "politikk, helse, sport, religion"
hypothesis_template: "Denne teksten handler om {}."
multi_class: true
After inspecting the actual request being sent by the browser to the HuggingFace server for the specific model, the parameter hypothesis_template was not even in the request payload. Is there a way to fix this? I guess that when specified in the model card, hypothesis_template should be a hidden input or just attached to the request JSON.
Cheers.