EasyInstruct: An Easy-to-Use Instruction Processing Framework for LLMs

Introduction

EasyInstruct is an open-source Python framework designed to simplify instruction processing for Large Language Models (LLMs). Accepted at ACL 2024, this project from zjunlp provides a modular and easy-to-use solution for researchers and developers. It streamlines the complex workflow of instruction generation, selection, and prompting, offering a comprehensive toolkit for experimenting with and enhancing LLM capabilities. EasyInstruct supports various state-of-the-art LLMs, including GPT-4, LLaMA, and Claude, making it a versatile choice for diverse NLP tasks.

Installation

Getting started with EasyInstruct is straightforward. You can install it directly from the GitHub repository or via PyPI.

Installation from git repo branch:

pip install git+https://github.com/zjunlp/EasyInstruct@main

Installation using PyPI (not the latest version):

pip install easyinstruct -i https://pypi.org/simple

For local development, clone the repository and install in editable mode:

git clone https://github.com/zjunlp/EasyInstruct
cd EasyInstruct
pip install -e .

Examples

EasyInstruct provides multiple ways to quickly get started, including shell scripts and a Gradio application. Here's a quick example using the SelfInstructGenerator to generate instruction data:

from easyinstruct import SelfInstructGenerator
from easyinstruct.utils.api import set_openai_key

# Step1: Set your own API-KEY
set_openai_key("YOUR-KEY")

# Step2: Declare a generator class
generator = SelfInstructGenerator(num_instructions_to_generate=10)

# Step3: Generate self-instruct data
generator.generate()

You can also explore the Gradio app for an interactive experience. Launch it locally or try the hosted version on HuggingFace Spaces:

python demo/app.py

Why Use EasyInstruct?

EasyInstruct stands out due to its modular design and comprehensive features for LLM instruction processing.

• Modular Architecture: It modularizes instruction generation, selection, and prompting, allowing for flexible combination and interaction of different techniques.
• Diverse Generation Methods: Supports popular instruction generation techniques like Self-Instruct, Evol-Instruct, Backtranslation, and KG2Instruct.
• Advanced Selection Metrics: Offers various metrics for selecting high-quality instruction datasets, including Length, Perplexity, ROUGE, GPT score, MTLD, and CIRS.
• Broad LLM Compatibility: Integrates with leading LLM API providers such as OpenAI (GPT-3.5, GPT-4), Anthropic (Claude), and Cohere (Command), along with local engine support.
• Ease of Use: Designed to be user-friendly, enabling researchers and developers to quickly set up and run experiments.
• Research-Oriented: Provides a robust framework for exploring and advancing research in LLM instruction tuning and prompt engineering.

Links

• Project Page: https://zjunlp.github.io/project/EasyInstruct
• Paper (arXiv): https://arxiv.org/abs/2402.03049
• HuggingFace Demo: https://huggingface.co/spaces/zjunlp/EasyInstruct
• Documentation: https://zjunlp.gitbook.io/easyinstruct/
• Video Overview: https://youtu.be/rfQOWYfziFo
• GitHub Repository: https://github.com/zjunlp/EasyInstruct