Prepare your model

What you supply

Two functions. Nothing gets rewritten:

def make_model():                              # a fresh, untrained model
    return LogisticRegression(max_iter=1000)

def load_data(partition_id, num_partitions):   # this client's local data
    ...
    return x_train, y_train, x_test, y_test

• scikit-learn (linear estimators) and PyTorch (nn.Module) are supported.
• The factory may be a plain function, a @staticmethod, or the nn.Module class itself. If it needs configuration, the wizard asks you for it — a config-driven model is normal, not an edge case.
• load_data may take (partition_id, num_partitions) or not. If it doesn't, every client loads the same thing unless your arguments say otherwise, and the wizard tells you so.

Analysis and validation

Upload a zip of your code. It is analyzed statically, never executed, and the generated adapter is validated in a locked-down sandbox (2 simulated clients, synthetic data, no network) before you get an artifact. A model that only looks right cannot ship.