Day 30

Today, I learned about prompt tuning.

The main model stays frozen.
↳The main model stays unchanged. In AI, it is called as frozen model.

Only a soft prompt is trained using examples from a specific task.
↳A soft prompt is not a readable sentence. It is a short sequence of vectors/groups of numbers added to the input.

later training adjusts these numbers so they guide the model toward a task, such as summarization/classification

Unlike prompt engineering, humans do not write a soft prompt word by word.

Also, Prompt Tuning is useful when you want to enhance the model's performance on specific prompts without requiring major retraining.

My takeaway:
Prompt tuning learns a small numerical instruction while leaving the main model unchanged.

Day 30 of 1% better.