In development

New features

Add TabICLv2 pre-training code (python -m tabicl.train): quantile regression training via a pinball loss (--regression_method quantile) in addition to classification, and the Muon optimizer (--muon True) alongside AdamW. The training CLI now also exposes the graph_scm prior options, layernorm-without-bias (--norm_type layernorm_nobias), SSMax (--col_ssmax/--icl_ssmax with --ssmax_type), feature grouping and target-aware embeddings (--col_feature_group, --col_target_aware, --col_affine), the RoPE variant (--row_rope_interleaved; v1 interleaved by default, v2 uses False), residual initialization (--zero_init; v2 uses False), and FlashAttention-3 during training (--use_flash_attn3; the v2 recipe enables it for stages 2 and 3 only). All CLI defaults reproduce the TabICLv1 model configuration; resuming a run re-seeds the data stream with the current step. Ships the three-stage TabICLv2 curriculum scripts, separately for the classifier and regressor checkpoints (scripts/train_v2_{clf,reg}_stage{1,2,3}.sh).

When unpickling a TabICL estimator, the fitted attributes device_, model_, etc. are only state if the pickled model was fitted. (PR#121)
Improve non-CUDA GPU inference reliability and performance (including XPU): inference now consistently runs on the configured backend device, uses backend-appropriate autocast, and queries available memory plus async stream/event primitives through backend-agnostic torch.<backend> APIs (with safe synchronous fallbacks when async is unavailable). This fixes pathological auto-batch sizing (e.g. batch size forced to 1) and restores expected accelerated inference behavior on supported non-CUDA GPU backends. When device=None, estimators now default to CUDA when available, otherwise XPU, then MPS, and then CPU. (PR#144)
Improve Apple Silicon MPS inference: MPS now uses the same AMP, auto-batching, and memory-aware inference path as other accelerators instead of falling back to the CPU path. use_amp="auto" is device-aware (off on CPU; size heuristic on CUDA/XPU/MPS), and float16 KV caches are kept on MPS when AMP is enabled. MPS is included in the default device order (CUDA → XPU → MPS → CPU). (PR#144)

Add SHAP and ShapIQ support with NaN-based feature masking to explain TabICL predictions, plus a dedicated tutorial and a faster SHAP path. (PR#90)
Add support for some unsupervised learning tasks. (PR#82)
Add support for raw quantiles (direct outputs of TabICL) in regression, enabling native quantile regression without post-hoc calibration. (PR#42)
Add preprocessing for NumPy array inputs, consistent with existing behavior for Pandas inputs: ordinal encoding for categorical features, mean imputation for numerical features, and encoding missing values as a separate category for categorical columns. (PR#51)

Clarify the public vs. private API boundary following scikit-learn conventions. Internal modules are now prefixed with an underscore (_model, _sklearn, _unsupervised, etc.); import public estimators (TabICLClassifier, TabICLRegressor, TabICLForecaster) from the top-level tabicl package. (PR#84)

Replace broadcasting with searchsorted in QuantileDistribution for faster quantile evaluation.

New documentation site built with Sphinx + Sphinx Gallery and published on Read the Docs, including a redesigned landing page. (PR#52, PR#54, PR#58, PR#60, PR#67, PR#81)
Add a project logo. (PR#74)
New tutorials: quantile regression (PR#61), probabilistic classification (PR#73), time series forecasting (PR#77), and skrub integration with string-handling fixes (PR#78).

Tweak dependency management and test against the development versions of dependencies in CI. (PR#53, PR#63)
Set explicit read permissions in CI workflows. (PR#76)

Drop Python 3.9 support and now requires Python >= 3.10
kv_cache moved from fit() to __init__() following scikit-learn convention. kv_cache is now a constructor parameter for both TabICLClassifier and TabICLRegressor.
TabICLForecaster API changes — output_selection renamed to point_estimate
Fix KV cache dtype mismatch. When AMP is enabled, cached projections stored in float16 caused errors when loaded on CPU/MPS/CUDA without AMP. The cache is now auto-upcast to float32 during loading.
Refactor time series forecasting module