TabICL in-context tabular-foundation discriminator (conditional single-sample)

Conditional-single-sample discriminator (family O) built on TabICL, the pretrained tabular foundation model of Qu et al. (2025). Like TabPFN, TabICL classifies a query row by in-context learning: a forward pass of a frozen pretrained transformer conditions on a provided labelled TRAINING context and emits the query's class posterior. There is no per-dataset gradient training – "fitting" only loads the training context into the model. This scorer is the direct sibling of fit_tabpfn, swapping the TabPFN forward calls for TabICL's.

Why this is single-sample-deployable (conditional-in-N). A specimen is single-sample-classifiable iff the TRAINING context that conditions the transformer is FROZEN into the model at fit time, so that at inference the query's posterior depends only on that query row and the frozen context – never on which other specimens are scored alongside it. This implementation freezes the per-sample-rCLR-transformed training context (rows + labels) and the feature universe at fit, and at score conditions every query against that one frozen context. The frozen context is the TRAINING rows only, so it is leakage-safe (no scored / test data ever enters the context).

Row-independent scoring. The deployment path evaluates predict_proba ONE QUERY ROW AT A TIME, so the \(n=1\) forward path is used uniformly and a query cannot be affected by the size or composition of a scoring batch. The legacy score_batch option is accepted for model compatibility but deliberately uses this same row-by-row path. GPU matrix kernels can otherwise produce small batch-size-dependent floating-point differences even when queries are conditionally independent.

Single-sample transform. Each specimen is mapped to the self-contained per-sample robust CLR (rCLR) over its OWN strictly-positive support (geometric-mean centring on v > 0) in the FROZEN feature universe; this is exactly invariant to per-specimen scaling and uses no cross-row statistic. The SAME rCLR transform is applied to the frozen training context at fit and to each query at score. Universe features absent from a specimen carry the neutral rCLR value 0 (centred log-ratio of an absent coordinate), matching the missing-feature rule of the other roster scorers.

Token-free pinned checkpoint. The default TabICL checkpoint (tabicl-classifier-v2-20260212.ckpt, the TabICLv2 model of Qu et al. 2025) is hosted in the public Hugging Face repository jingang/TabICL and is fetched non-interactively by hf_hub_download – there is no license acceptance or interactive token step (unlike TabPFN's gated V2.6). This method pins that checkpoint version explicitly so the model is reproducible and the download path is token-free.

Orientation and output. The score is the frozen-context posterior \(P(\mathrm{case}) = P(y = 1)\) for the query, in \([0, 1]\); larger = more case-like. Degenerate queries (fewer than hp$min_features universe features present, or an all-zero / empty specimen) return the neutral probability 0.5.

References

Qu J, Holzmuller D, Varoquaux G, Le Morvan M. (2025) TabICL: A Tabular Foundation Model for In-Context Learning on Large Data. International Conference on Machine Learning (ICML). arXiv:2502.05564.

Hollmann N, Muller S, Purucker L, et al. (2025) Accurate predictions on small data with a tabular foundation model. Nature 637:319-326.