The task is to classify email text bodies into exclusive categories like feedback, complaint etc. I have a labelled dataset available having about 350 samples.

I have tried the facebook/bart-large-mnli zero shot classification model where I passed the class names as possible label. It is already giving a decent performance.

Now, if I want to improve by using the existing labelled dataset and some model like distilbert-base-uncased, then will I lose the capability offered by the zero shot model altogether, and the new model will be trained entirely based on the labelled data?

I am afraid to go down the route because the number of labelled samples is so small, I feel it will fail to update the huge models having more weights than number of samples (we know the larger the model, the more samples you need).

So how do you guys address this concern, and how best to use the capabilities of the zero shot model on huggingface, while also using the labelled sample somehow?

I feel if I could just nudge the zero shot model a bit with the training samples, that would be the best possibility, but how to achieve it with huggingface?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.