Fine tuning a model for classification feels similar to machine learning using test and training data. Instruction fine tuning is also a type of fine tuning but it requires bigger data and more resources basically classification -> pretty specialized instruction -> a bit general but instructionx.
Basically in this chapter we break down the outer layers of llm to fine tune we use a dataset with spam and ham sms.
We undersample the data to make sure we have similar numbers for each label to create a balanced dataset.
Same as ml we split into training validation and test data, training 70 validation 10 and test 20%
Creating Data Loaders
Uniformly Sized Text Chunks (Sliding Window) -> Batching Same question here lose data by truncating to the smallest or padding messages to the longest message to get everything in order to not lose data.