2. Create a Task

📍 Where you are: Task Owner 2/7 — ① Design → ② Create Task → ③ Workspace → ④ Train · publish → ⑤ Approve → ⑥ Campaign → ⑦ Results
Screen: FedOps Console · Next: open this Task’s Workspace in Studio

🎯 What you do in this step: Create an empty shell (Web Draft) in FedOps saying “this is the federated learning I want to run”. You decide the name, what kind of model it is, and who can see and join it. There is no code, data, or model yet; you fill those in Studio in the next chapter. Nothing is published yet at this step.

  1. 🖱️ Follow along
  2. ⚙️ Options explained
    1. Name and identity
    2. Kind of model
    3. Visibility and participation — the two most important fields
    4. Description
    5. What you do not decide here
  3. ✅ Checklist
  4. Next

🖱️ Follow along

Create Federated Task button on the My Federated Tasks screen

  1. In the FedOps Console top menu My Federated Tasks, click Create Federated Task.

Entering the Task name handson-heart and the Data modality

  1. Enter handson-heart in Federated Task name. Registry ID below fills in automatically.
  2. Change Data modality to Tabular.
  3. Leave the rest as is and click Create. (See the options explained below for each field.)

The created handson-heart Task management screen

Success when the Task management screen titled handson-heart appears. The tabs at the top (Registry Release, Participants, Server Management, Monitoring, Global Model) are used in the following chapters.

⚙️ Options explained

Name and identity

FieldWhat it decidesThis tutorial
Creation ModeFederated Task (FedOps 1.3) or Legacy (FedOps 1.2 style)1.3 — Legacy only to continue an existing 1.2 Task
Federated Task nameHuman-readable name. Becomes the Studio Workspace name and later the Agent’s Tool id (local:handson-heart)handson-heart
Registry IDID that distinguishes the Task in the Registry. Once published it becomes @my-handle/registry-idAutomatic from the name
Primary Model working nameTemporary model name. The final name shown in the Registry comes from model.display_name in the code’s config.yamlLeft empty — final name is the template’s Heart Disease MLP

Kind of model

FieldChoicesIf you change itThis tutorial
Model roleAI · Tool AI / LLM · Base LLMChanges where Agent Builder suggests the finished model. Tool AI is a tool that outputs numbers (predictions); Base LLM is the language model that writes sentencesAI · Tool AI — our model is a classifier that outputs probabilities
Task categoryClassification / Regression / Generation / Representation learning / OtherCategory shown in the Registry. The training itself is decided by the codeClassification — two classes, no disease / disease
Data modalityUndecided / Image / Tabular / Time series / Text / MultimodalShown in the Registry and used for searchTabular — a table of 10 test values per patient

Changing Task category or Data modality does not change the training code. It is information Clients use in the Registry to judge “what this Task does”.

Visibility and participation — the two most important fields

FieldChoiceWhat changes
VisibilityPrivateOnly authorized people see the Task. It is not listed in the Registry
 PublicListed in the Registry. Task information and downloads of registered Global Models become public. Server control, YAML, logs, and Owner functions are not public
ParticipationOwner approval requiredA Client must request to join and the Owner must approve
 Open (automatic approval)Requests are approved immediately

This tutorial’s choice:

  • Start with Visibility = Private. There is no reason to show an empty Task before code and model are ready. It becomes Public when you publish to the Registry in chapter 4.
  • Participation = Owner approval required. In medical federated learning, the Owner needs to know “who is adding data to our model’s training”. If anyone could join and train on bad data, the Global Model could break. It also lets you experience the request → approval process in this tutorial.
📚 Federated learning background — why screen participants

A federated learning server averages the parameters clients send on trust. It cannot see the data, so verification is hard. Therefore:

  • Data quality issues: a single client trained on badly labeled data can make the Global Model worse.
  • Malicious participation (poisoning): attacks that send parameters designed to break the model or make it misbehave on specific inputs have been studied.

Real medical federated learning usually fixes the participating hospitals by contract in advance (cross-silo). Think of the Hungarian and Cleveland hospitals in this tutorial as having that relationship. Owner approval required brings that practice onto the screen. Open is for lowering the barrier, as in open research collaborations.

Description

FieldWhat it decides
Public summaryOne-line description on the Registry card (280 characters). If Private, only authorized people see it
Federated Task tagsTags for search (comma separated)

The detailed description Clients actually read is not this field but the code’s README.md. The README is shown as is on the Registry Task card, and Clients check the data format there (10 columns and target). You will see it in chapter 4.

What you do not decide here

The number of rounds, number of participants, and aggregation method are not decided when creating the Task. You decide them each time you train, in the campaign in chapter 6. The model structure, preprocessing, and training settings are decided in the code (config.yaml) in the next chapter.

✅ Checklist

  • handson-heart is in My Federated Tasks in the Console
  • You can explain what you chose for Visibility and Participation, and why

Next

3. Put Code and Data into the Workspace — open the empty Task in Studio and fill in the code every participant will run.


Back to top

© 2026 Cognitive Computing Lab, Gachon University