Fine Tuning for Instruction

Main idea

Train the base model on schema that includes prompt and appropriate response. Example:

           Instruction: Tell me the capital of Ohio.
           Response: Columbus is the capital of Ohio.
        

Considerations

  • The model architecture does not change.
  • The schemas need to be batched for effective training.
  • The training is more extensive than that required for classification and may need a larger model.
  • The generated response requires post-processing.

Raschka demo

Zipped folder with results from medium-sized GPT model