Mixed-data preprocessing

With preprocess="auto", DataFrame dtypes determine the processing path.

• Numeric columns: median imputation, then standardization.
• Other columns: most-frequent imputation, then dense one-hot encoding.
• Unknown test categories: ignored rather than treated as errors.
• All preprocessing: fitted on the training fold only.

model = SurvFMRMSTRegressor(
    backbone="tabpfn",
    tau=365,
    preprocess="auto",
)
model.fit(X_train_dataframe, time_train, event_train)

For an already encoded numeric matrix, automatic preprocessing passes values through after dimensionality and finiteness checks.

Custom preprocessing

Pass any transformer that implements fit_transform(X) and transform(X):

model = SurvFMRMSTRegressor(
    backbone="linear",
    tau=365,
    preprocess=my_pipeline,
)

Preprocessing may change the processed feature count. This matters for hosted services with column limits; inspect model.get_metadata() after fitting.