Week 2 - PEFT/Soft prompts - difference between full and multi-task fine-tuning

Course Q&A Generative AI with Large Language Models
1

In Week 2, lecture “PEFT techniques 2: Soft prompts”, a chart is shown to illustrate how well soft prompting performs for large models.

image
image2774×616 115 KB

The chart includes “Full fine-tuning” model and “Multi-task fine-tuning” model.

I have a question about the difference between these two. I thought that full fine-tuning really is multi-task. Here, does “Full fine-tuning” refer to single-task full fine-tuning?

If yes, is that the one discussed in earlier lectures about catastrophic forgetting?

What is meant by “Multi-task fine-tuning” in terms of the previously covered lectures? Is it full fine-tuning trained by multi-task instruction prompts? (like FLAN-T5 or FLAN-PALM)

thanks

2

Full fine-tuning means fine-tuning the model (updating all of its parameters) on a particular task and multi-task fine-tuning on several tasks that are probably similar to each other. In these type of tunings the models are not frozen and all the parameters are updated in the process.

3

Yes but the model is not frozen!