Backbone setup

All registered backbones solve the same pseudo-RMST regression problem. Their dependencies, access requirements and persistence behavior differ.

Name	Execution	Install	Important setup
`tabpfn`	Local	`.[tabpfn]`	Accept the model license and configure checkpoint access.
`tabicl`	Local	`.[tabicl]`	First use may download a checkpoint.
`tabdpt`	Local	`.[tabdpt]`	Prediction uses `n_ensembles=8` and seed `20260715` by default.
`tabh2o`	Hosted API	`.[tabh2o]`	Set `TABH2O_API_KEY`; data leave the local machine.
`mitra`	Local	`.[mitra]`	Separate AutoGluon environment recommended; direct `Mitra` output only.
`linear`	Local	`.[sklearn]`	Lightweight diagnostic baseline.
`random_forest`	Local	`.[sklearn]`	General local starting point without model access.
`gradient_boosting`	Local	`.[sklearn]`	General local baseline.

TabPFN

Follow the official TabPFN quickstart and model-access guide. Do not place access tokens in source files or notebooks committed to version control.

model = SurvFMRMSTRegressor(
    backbone="tabpfn",
    backbone_kwargs={"device": "cuda", "model_path": "/path/to/checkpoint"},
    tau=365,
)

TabICL and TabDPT

Use the official TabICL repository and official TabDPT repository for installation and checkpoint instructions.

TabH2O

export TABH2O_API_KEY="..."

model = SurvFMRMSTRegressor(backbone="tabh2o", tau=365)

The adapter reads the key only when predict() sends a request. The credential value is not stored on the estimator, in its metadata or in saved model files. Feature names are anonymized in the payload. The hosted service receives the processed training table, pseudo-targets and test features; do not use it for protected or restricted data without institutional approval.

MITRA

The adapter follows the official AutoGluon foundation-model workflow. It requires the direct Mitra base model and does not silently substitute an AutoGluon weighted ensemble. Use AutoGluon's native artifact management.