ML Engineering

TabICL

TabICL and the tabular in-context foundation-model family evaluated by the TabArena living benchmark, ingested at the GRANULAR per-row level: each response is one tabular foundation model's per-row correctness on one individual OpenML classification test row (fold 0 of the tabarena-v0.1 suite). The response is per-row classification correctness (argmax of the predicted class probabilities vs the true label). Items are real tabular rows (feature name->value). Subjects are in-context tabular foundation models: TabICL, TabPFNv2, TabDPT, TabFlex, Mitra, BetaTabPFN, LimiX.