Task Configuration¶

When creating a new project, the Tasks field requires a JSON array that defines the federated learning workflow.

For the complete task configuration reference with all model-specific options, see the full TASK_GUIDE.

Task Structure¶

Each task must have:

seq: Sequential number (starting from 1)
model: The ML model class name
config: Configuration dictionary

[
  {
    "seq": 1,
    "model": "LogisticRegression",
    "config": {
      "total_round": 5,
      "current_round": 1
    }
  }
]

Available Models¶

Classification & Regression¶

Model Name	Description	Dataset Format
`LogisticRegression`	Binary classification	Features + binary label (0/1)
`RLogisticRegression`	R version of logistic regression	Same
`LogisticRegressionStats`	Statistical logistic with inference	Features + binary label, min 30 samples
`LinearRegression`	Continuous value prediction	Features + continuous target
`SvmRegression`	Support Vector Machine regression	Features + continuous target
`Ancova`	Analysis of Covariance	Groups + covariates + outcome
`OrdinalLogisticRegression`	Ordered categorical outcomes	Features + ordinal label (0,1,2,...)
`MixedEffectsLogisticRegression`	Clustered binary data	Group ID + features + binary label

Survival Analysis & Censored Outcomes¶

Model Name	Description	Dataset Format
`CoxProportionalHazards`	Time-to-event regression	Features + time + event (0/1)
`RCoxProportionalHazards`	R version (survival::coxph)	Same
`KaplanMeier`	Non-parametric survival estimation	Group + features + time + event
`RKaplanMeier`	R version (survival::survfit)	Same
`CensoredRegression`	Tobit Type I (left/right censoring)	Features + outcome + censoring (-1/0/1)
`RCensoredRegression`	R version (survival::survreg)	Same

Count Data Models¶

Model Name	Description	Dataset Format
`PoissonRegression`	Count data with rate ratios	Features + offset + count
`RPoissonRegression`	R version (glm family=poisson)	Same
`NegativeBinomialRegression`	Overdispersed count data	Features + offset + count
`RNegativeBinomialRegression`	R version (MASS::glm.nb)	Same

Missing Data¶

Model Name	Description	Dataset Format
`MultipleImputation`	MICE with Rubin's rules	Features (may have NaN) + outcome
`RMultipleImputation`	R version (mice::mice)	Same

Image Segmentation¶

Model Name	Description	Dataset Format
`FederatedUNet`	UNet with FedAvg aggregation	Zip of `images/` + `masks/` directories

Required Config Parameters¶

Parameter	Description
`total_round`	Total number of federated learning rounds
`current_round`	Starting round (usually 1)

Optional Config Parameters¶

Parameter	Applies To	Description
`local_epochs`	Classification/Regression	Local training epochs per round
`learning_rate`	Classification/Regression	Training learning rate
`n_group_columns`	ANCOVA	Number of group indicator columns
`vcp_p`	Mixed Effects Logistic	Prior SD for variance components (default: 1.0)
`fe_p`	Mixed Effects Logistic	Prior SD for fixed effects (default: 2.0)
`m`	Multiple Imputation	Number of imputed datasets (default: 5)
`max_iter`	Multiple Imputation	Max MICE iterations (default: 10)
`patch_size`	FederatedUNet	Image resize dimension (default: 64)
`architecture`	FederatedUNet	Encoder backbone, e.g. `resnet50`, `mobilenetv2` (default: resnet50)
`type_Unet`	FederatedUNet	Segmentation architecture type (default: unet)
`batch_size`	FederatedUNet	Training mini-batch size (default: 8)

Example Configurations¶

Logistic RegressionCox PH (R)Censored RegressionMICEFederatedUNet

[{"seq": 1, "model": "LogisticRegression", "config": {"total_round": 5, "current_round": 1}}]

[{"seq": 1, "model": "RCoxProportionalHazards", "config": {"total_round": 1, "current_round": 1}}]

[{"seq": 1, "model": "CensoredRegression", "config": {"total_round": 1, "current_round": 1}}]

[{"seq": 1, "model": "MultipleImputation", "config": {"total_round": 1, "current_round": 1, "m": 5, "max_iter": 10}}]

[{"seq": 1, "model": "FederatedUNet", "config": {"total_round": 1, "current_round": 1, "local_epochs": 1, "architecture": "resnet50", "type_Unet": "unet", "patch_size": 64, "batch_size": 1, "learning_rate": 0.0001}}]

Agent Hooks (Optional)¶

Add an agent block to the task config to enable LLM-powered hooks during training:

[{
  "seq": 1,
  "model": "LogisticRegression",
  "config": {
    "total_round": 10,
    "current_round": 1,
    "agent": {
      "enabled": true,
      "summaries": true,
      "early_stopping": true,
      "outlier_detection": true
    }
  }
}]

Agent Parameter	Description
`enabled`	Master switch for all agent hooks
`summaries`	Generate natural-language summaries after each training round
`early_stopping`	Evaluate convergence after aggregation; stop early if model has converged
`outlier_detection`	Compare cross-site artifacts before aggregation; flag anomalous sites

Requires ANTHROPIC_API_KEY environment variable and poetry install --extras agent. Without these, all hooks are no-ops and training proceeds normally.

Writing R-Based Tasks¶

R tasks use a Python-R bridge via AbstractRTask. Each R task needs:

A Python wrapper class that sets r_script_dir
Three R scripts in a scripts/ subdirectory:
- prepare_data.R -- validate data
- training.R -- fit model
- aggregate.R -- meta-analyze across sites

See Architecture > Adding a New ML Task for details.