The process of retraining a pre-trained model’s parameters based on a specific often smaller dataset

SFT - Supervised Fine Tuning. The dataset is provided explicitly telling what data is beforehand

We’re not creating a new model, we’re just creating a new outputted state of the model

Types of fine tuning

Changing the dataset

  • Instruction fine tuning - good and bad prompt examples provided
  • Domain specific fine tuning - a new knowledge base to be provided (blogs, docs) Changing the method of training
  • Full Fine Tuning - All of the models weights are updated, and expensive
  • PEFT - Parameter efficient fine tuning, update small set of parameters and freezes the rest
  • Last Layer Fine Tuning - Freeze all the layer except the last layer Removing Params and Weights
  • Pruning - Removing parameters, to make the model smaller and efficient (less compute)
    • Train-Time Pruning
    • Post-Training Pruning