PEFT (Prompt Engineering with Fine-Tuning)

MindNLP's PEFT (Prompt Engineering with Fine-Tuning) is a methodology for fine-tuning natural language models using prompts. It allows users to tailor their models for specific tasks or domains by providing customized prompts, enabling better performance on downstream tasks.

Introduction

PEFT leverages the power of prompts, which are short natural language descriptions of the task, to guide the model's understanding and behavior. By fine-tuning with prompts, users can steer the model's attention and reasoning towards relevant information, improving its performance on targeted tasks.

Supported PEFT Algorithms

Algorithm	Description
AdaLoRA	Adaptable Prompting with Learned Rationales (AdaLoRA)
Adaption_Prompt	Adaptation Prompting
IA3	Iterative Alignments for Adaptable Alignment and Prompting (IA3)
LoKr	Large-scale k-shot Knowledge Representation (LoKr)
LoRA	Learnable Redirection of Attention (LoRA)
Prompt Tuning	Prompt Tuning fine-tunes models by optimizing the prompts used during fine-tuning.

Each algorithm offers unique approaches to prompt engineering and fine-tuning, allowing users to adapt models to diverse tasks and domains effectively.