Prefix Tuning Architecture
Prefix tuning is a parameter-efficient fine-tuning method where a sequence of trainable vectors, known as prefixes, are prepended to the hidden states at each layer of a Transformer model. For any given layer l, the input consists of the prefixes for that layer (e.g., ) followed by the hidden states () computed by the previous layer. The core LLM parameters remain frozen, while only these layer-specific prefix vectors are optimized during training to steer the model's output for a downstream task.
0
1
Tags
Ch.3 Prompting - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Related
Input Composition Formula for Prompt Tuning
An engineer is adapting a large, pre-trained language model for a new task. To do this efficiently, they keep all the original model's parameters frozen. Their adaptation strategy involves modifying the input sequence before it is processed by the model. For any given text, they first convert the text into its standard sequence of numerical token representations. Then, they prepend a separate, short sequence of newly initialized, trainable numerical vectors to the beginning of that sequence. Only these new vectors are updated during training on the new task. Which statement best distinguishes the nature of these prepended, trainable vectors from the standard token representations?
You are examining the input layer of a large language model adapted using a parameter-efficient technique. The input is formed by combining two distinct types of numerical vectors. Match each vector type with its correct description.
Prefix Tuning Architecture
Parameter-Efficient Model Adaptation
Learn After
A researcher is implementing a parameter-efficient fine-tuning method for a large language model. The goal is to adapt the model to a new task by introducing a small number of new, trainable parameters while keeping the vast majority of the original model's weights frozen. Which of the following implementation strategies correctly identifies the unique architectural modification central to this specific method?
In the prefix tuning architecture, a sequence of trainable vectors is prepended exclusively to the initial input embeddings of the model. The hidden states of all subsequent layers are then computed based on this modified input, without any further addition of trainable vectors at those deeper layers.
Analyzing an Implementation of a Fine-Tuning Method