flwr new @gfedops/flowertune-gccl-medical-llm

FlowerTune GCCL Medical LLM

Federated instruction tuning with ContactDoctor/Bio-Medical-Llama-3-8B on medalpaca/medical_meadow_medical_flashcards using Flower Simulation Engine.

Quickstart

pip install -e .
huggingface-cli login
flwr run .

Project Structure

flowertune-gccl-medical-llm/
├── flowertune_gccl_medical_llm/
│   ├── __init__.py
│   ├── client_app.py
│   ├── server_app.py
│   ├── dataset.py
│   ├── models.py
│   └── strategy.py
├── pyproject.toml
└── README.md

Configuration

All app settings are defined in pyproject.toml under [tool.flwr.app.config].

Key defaults:

• model.name = "ContactDoctor/Bio-Medical-Llama-3-8B"
• model.quantization = 4 (bitsandbytes 4-bit)
• model.lora.peft-lora-r = 32
• model.lora.peft-lora-alpha = 64
• strategy.fraction-fit = 0.1
• num-server-rounds = 100
• dataset.name = "medalpaca/medical_meadow_medical_flashcards"

You can override run-time values:

flwr run . --run-config "num-server-rounds=10 strategy.fraction-fit=0.2"

Method

• PEFT LoRA fine-tuning via peft + trl (SFTTrainer)
• FedAvg-based aggregation strategy
• Cosine annealing LR schedule between:
  • train.learning-rate-max
  • train.learning-rate-min
• IID partitioning with flwr-datasets (num-supernodes clients)

Outputs

• Global PEFT checkpoints are saved on server rounds:
  • every train.save-every-round
  • and final round
• Saved path format:
  • results/<timestamp>/peft_<round>
• Communication budget tracking is appended to:
  • comm_cost.txt

Compatibility Notes

This app was validated with:

• flwr==1.26.1
• torch==2.3.1
• transformers==4.43.1
• trl==0.8.1
• bitsandbytes==0.43.3
• accelerate==0.31.0

If DataCollatorForCompletionOnlyLM is missing in your trl version, the app falls back to default collator behavior.

Flower Hub

After publishing, users can fetch this app via:

flwr new @gfedops/flowertune-gccl-medical-llm