ML//prompt tuning

Prompt tuning is a parameter-efficient method of adapting a frozen model to a task by training a small set of continuous *soft prompt* vectors prepended to its input.

Prompt tuning is a parameter-efficient method of adapting a frozen model to a task by training a small set of continuous soft prompt vectors prepended to its input.

It is much cheaper than fine-tuning, because only the prompt vectors are trainable (a few thousand parameters), and a learned prompt embedding can outperform hand-written prompts, especially on classification tasks: the goal of prompt engineering is kept, steering the model, but the prompt becomes continuous vectors in place of discrete tokens. Prefix-tuning places soft prompts at every layer; prompt tuning proper (Lester et al.) places them only at the input embedding layer.