FedOps Tutorial
Learn both roles of federated learning — Task Owner and Client — hands-on with one public medical dataset, in FedOps 1.3
This tutorial does not just list buttons to click. At every step it explains what you are doing, why, and what changes if you pick a different option, together with the federated learning background behind it. The goal is that, once you finish, you can apply the same flow to your own data and your own model.
What you will do
Two hospitals in different countries train a heart disease classifier (no disease / disease) together without pooling their patient records. The data is UCI Heart Disease, which was actually collected and published hospital by hospital — the Task Owner holds the records of a hospital in Hungary, and the Client holds the records of the Cleveland hospital in the US. The Owner designs and runs the training; the Client joins with its own data. When training is done, you attach the model to an LLM Agent as a tool and call it over an API.
Pick your role first
| I want to… | Reading order |
|---|---|
| Open and run federated learning (Task Owner) | Basics → Task Owner Track → (optional) Model Use Track |
| Join someone else’s federated learning with my data (Client) | Basics → Client Track |
| Try both roles on my own | Basics → Owner Track 1–5 → Client Track → Owner Track 6–7. One account can be both Owner and Client |
| Plug the finished model into my service | Basics → Model Use Track |
The whole flow — where the two roles meet
sequenceDiagram
autonumber
participant O as Task Owner<br/>(Hungarian hospital)
participant F as FedOps<br/>(Console · Registry · aggregation server)
participant C as Client<br/>(Cleveland hospital)
O->>O: Design data and model (without FedOps)
O->>F: Create Task · publish code and Initial Model
C->>F: Find the Task in the Registry · request to join
F-->>O: Join request notice
O->>F: Approve the participant
C->>C: Open Workspace · place my data · readiness check
O->>F: Configure campaign · start server
loop Every round
F->>C: Current Global Model parameters
C->>C: Local training on my data
C->>F: Trained parameters only (no data is sent)
F->>F: Merge with FedAvg
end
F-->>O: Global Model vN · metrics
F-->>C: My participation record · metrics
The two roles meet in only three places: ③ join request ↔ ⑤ approval, ⑦ server start ↔ joining rounds, and the results. Everything else each side does on its own screen. The 📍 Where you are box at the top of every page tells you where you are in this picture.
How to read the pages
Every step page follows the same frame.
| Box | Content |
|---|---|
| 📍 Where you are | Your current position in the whole flow and the next step |
| 🎯 What you do in this step | What you do and why |
| 🖱️ Follow along | Real screenshots. Red boxes + numbers are where to click, and match the numbers in the text |
| ⚙️ Options explained | What each option on the screen decides, what changes if you change it, and why this tutorial picks that value |
| 📚 Federated learning background | The federated learning concept behind the choice (collapsed box) |
| ✅ Checklist | Fill it in, then move on |
What you need and how long it takes
| Item | Requirement |
|---|---|
| PC | macOS / Windows / Linux, 16 GB memory recommended, 20 GB+ free disk |
| Docker | Latest Docker Desktop (or Docker Engine + compose) |
| Python | 3.10 or later, below 3.13 (3.12 recommended) |
| Accounts | FedOps accounts. This tutorial assumes one Owner + one Client, two accounts (two PCs recommended) |
| Data | The practice template contains the original UCI Heart Disease files. prepare_data.py builds per-hospital CSVs; the Owner uses hungarian (Hungary) and the Client uses cleveland (Cleveland) |
| Track | Time |
|---|---|
| Basics (including install and LLM download) | 1–2 hours |
| Task Owner | 2–3 hours |
| Client | 1 hour |
| Model Use | 1–2 hours |
Good to know
- This is public data for education. Do not use the model built here for diagnosis.
- The goal of this tutorial is not high accuracy but running federated learning end to end with FedOps. Accuracy on this data is around 0.75–0.8, and the numbers on screen are measured examples.
- Each participant uses a different hospital’s data. During federated learning nobody sends or sees anyone else’s data.
- Screens are as of FedOps 1.3 and Agent Studio, October 2026. Button names may change in later versions.
- Most screens are in English. Some messages from the practice template and the demo app are in Korean; this guide shows them as they appear, with an English translation in parentheses.
- Practice template: https://github.com/gachon-CCLab/FedOps-hands-on-guide
- Data source: Janosi, Steinbrunn, Pfisterer, Detrano (1989), Heart Disease, UCI Machine Learning Repository, https://doi.org/10.24432/C52P4X (CC BY 4.0)