Fine-tuning NLP transformers for task automation - Jon Eskin's Blog

This article highlights what I suspect is going to be a very common and powerful way to get benefit from AI. There are many tasks that need to be done where an "intelligent" approach would be a big win, but you don't need "too much" intelligence. Why not bring a simple model with very defined training to the problem.

For simplicity, I ran with the assumption that each string of text had one and only one classification. It would be much more useful if each string of text instead could have 0-3 classifications. This would also almost certainly bring down the accuracy of the model because determining whether a given string belongs to a single category at all seems like it would be a much more complex task than what it had to learn during training.

After 3 epochs, the model achieves an accuracy of 100% and loss of 0.156 on the test set and an accuracy of 99.59% and loss of 0.0278 on the training set.

Here the test was to look at emails and classify them. There are many, many tasks like this in modern workflows.