What is the maximum number of predictable classes and how to do fine tune?🤗
hello everyone i would like to ask what is the maximum number of classes predictable by this model and more importantly what is the confidence with which it predicts the Italian language? however having a very high number of classes in the order of hundreds what advice could you give me? is it the case to do fine tune for such a high number of classes? if so how can i do it?
thank you all!!🤗
Technically, there is no limit for the classes you can predict with NLI zero-shooting. Performance will, however, decrease as you add more classes. Hard to give a concrete number, because it strongly depends on the complexity of your classes. So hundreds of classes works, but will not perform very well (that's quite a lot for any classifier).
Regarding Italian: Italian was both in the pre-training corpus of mDeBERTa as well as the NLI fine-tuning corpus, so performance on Italian should be good.
Fine-tuning an NLI model for other tasks than NLI takes some experience. I'd recommend you look into the Hugging Face course for the fundamentals. Then you can find more details in this paper and also in the google colab notebook in footnote 2: https://osf.io/xc39s