Glossary term
Glossary term
Training and Fine-Tuning
PEFT technique that learns a small set of soft prompt tokens prepended to input embeddings, with all model weights frozen.
A parameter efficient tuning mechanism that learns a "prefix" that the system prepends to the actual prompt.
One variation of prompt tuning—sometimes called prefix tuning—is to prepend the prefix at every layer. In contrast, most prompt tuning only adds a prefix to the input layer.
Click the icon to learn more about prefixes.
Lester et al. (2021, Google) showed that prompt tuning matches full model fine-tuning performance for T5-11B while training only 20K parameters vs 11B - demonstrating that at sufficient model scale, soft prompts are as effective as weight updates.
Google uses prompt tuning to create task-specific variants of large T5 and PaLM models for internal applications, maintaining a single set of base model weights while serving hundreds of tasks via task-specific prompt embeddings.
Multitask Prompt Tuning (MPT, 2023) learns a shared prompt that can be composed with task-specific prompts, enabling few-shot generalisation to new tasks with 2.5M parameters vs 100M for adapter-based methods.
Created for this library
A SaaS team uses prompt tuning to specialize a base LLM per customer without fine-tuning the full model.
A research team uses prompt tuning to test many domain adaptations of a foundation model with a small parameter footprint.
An enterprise legal team uses prompt tuning to specialize a base model on contract review while keeping the base weights shared.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License