Select your model and dataset, configure FL training, and download a ready-to-run federated VLM fine-tuning setup. Create a task at ccl.gachon.ac.kr/fedops/task, then deploy the same folder on server (K8s) and client (local GPU).
Vision Language Model to federate
LLaVA-OneVision with Qwen2 0.5B backbone. Fast training, low memory. Includes pre-generated parameter_shapes.json.
MLC-compatible VLM optimized for Samsung Galaxy A24 edge deployment. Larger model, richer representations.
Federated dataset for VQA fine-tuning
Medical radiology visual QA — 313 training samples. Validated with PhiVA (39.2% exact match).
General visual QA benchmark — large scale, diverse image-question pairs across many domains.
Pathology visual QA — histology and microscopy images with yes/no and open-ended clinical questions (~19.7K train).
Multi-organ medical VQA — brain, chest, abdomen across X-ray, CT, MRI. English subset (~4.9K train, 1K test).
Chart and figure visual QA — reasoning over bar charts, line graphs, and pie charts (~18K train).
Document visual QA — scanned business and industrial documents with information extraction questions (~10K train).
Federated learning hyperparameters
Total aggregation rounds
Gradient steps per FL round
Higher = more params, more VRAM
4-bit recommended for T4 GPU
Clients required per round
FedMAP solves a QP each round to minimise directional variance — better for cross-domain non-IID
AdamW learning rate
Generated config.yaml and zip contents
# Select model and dataset above to generate config
FedOps-Multimodal-Setup.zipbash setup.shpython server_main.py:8080python client_main.pypython client_manager_main.pyFEDOPS_PARTITION_ID=N