I used “facebook/bart-large-mnli” model for zero-shot classification.
I have a dataset of 3k entries and 27 labels. I split it 80:20 for train/ test. I fine-tuned the model on this dataset.
The NLI model performed very well in the test data, giving me a 95% accuracy. But when I used this fine-tuned model for Zero shot classification, the accuracy dropped to 80%. Is there any way I can increase the ‘classification accuracy’?
NOTE: One way is to probably change the labels. I want to know if there is any other way to do this.
PS: I use HuggingFace Trainer to train the model.