MNLI zero-shot-classifier model card

The pipeline for the zero-shot-classifiers is really awesome, and the demo is impressive. However, the demo does not seem to accept a hypothesis_template. I am publishing a Norwegian zero-shot-model, and when using the English template, the results are really bad.

I have tried specifying the hypothesis_template in the model card, but it does not seem to respect this.

Any ideas how to fix this?