We believe in a world where the default AI path for developers is trustworthy, safe, and open
Learn more
Last updated
3/10/2025
Get started
Federated AI
LLM
Finetune an LLM using Federated AI
This Blueprint demonstrates how to finetune LLMs using Flower, a framework for federated AI. As access to high-quality public datasets declines, federated AI enables multiple data owners to collaboratively fine-tune models without sharing raw data, preserving privacy while leveraging distributed datasets.
We apply parameter-efficient fine-tuning (PEFT) with LoRA adapters to fine-tune the Qwen2-0.5B-Instruct model on the Alpaca-GPT4 dataset. This approach optimizes resource efficiency while maintaining model adaptability, making it a practical solution for decentralized AI development.
Flower
Time
20 min
Complexity
High
Requires significant expertise
Hosted demo
Step by step walkthrough
Tools used to create
Trusted open source tools used for this Blueprint
HuggingFace Transformers
HuggingFace Transformers is used for model finetuning and HuggingFace Datasets is used for loading the dataset.
Choices
Insights into our motivations and key technical decisions throughout the development process.
Focus
Decision
Rationale
Alternatives Considered
Trade-offs
Federated AI Framework
Federated AI Framework
Used Flower for federated fine-tuning.
Developed by Flower.ai team, with the required expertise to create the Blueprint.
Fine-tune the model locally.
Will require one to gather huge amounts of data, pay for licenses, which will be a costly solution, and performance decreases.
Base Model
Base Model
Fine-tuned Qwen2-0.5B-Instruct.
Smaller size to make federated fine-tuning more accessible for initial experimentation.
Larger/even smaller models from different models series (Qwen, Llama, etc.).
Larger models require more compute; even smaller models may lose expressiveness.
Simulation vs. Deployment
Simulation vs. Deployment
Default simulation mode for federated training.
Easier for developers to test without extensive infra setup.
Direct deployment with Flower’s Deployment Engine.
Simulations may not capture all real-world constraints.
Training Hardware
Training Hardware
Supports CPU and GPU fine-tuning.
Increases accessibility for users with limited compute resources.
GPU-only training for efficiency.
CPU training is significantly slower, especially for larger models.
Dataset
Dataset
Used Alpaca-GPT4 for fine-tuning.
Well-structured dataset for instruction tuning (not too large).
Custom datasets.
Alpaca-GPT4 may not cover edge cases for specific use-case.
Demo and Evaluation
Demo and Evaluation
Provided both Streamlit app and CLI-based evaluation for interactive testing.
Simple way to validate model responses in real time, depending on user-preference.
Real-world deployment across the globe. It's feasible.
Would need to find partners to set this up or rent more instances across the world.
Ready? Try it yourself!
System Requirments
OS: Linux Python 3.10 or higher Minimum RAM: 8GB (recommended for LLM fine-tuning)
Learn More
Help Documentation
Detailed guidance on GitHub walking you through this project installation.
View More
Discussion Points
Get involved in improving the Blueprint by visiting the GitHub Blueprint issues.
Join in
Explore Blueprints Extensions
See examples of extended blueprints unlocking new capabilities and adjusted configurations enabling tailored solutions—or try it yourself.
Load more
