Causilo gives data scientists a ready‑to‑use tabular foundation model that works out of the box for both classification and regression. The main pain points teams face are: needing a model that handles mixed data types and missing values without heavy preprocessing, wanting fast inference on modest hardware, and requiring a clear path from research to production without legal headaches.
Causilo solves these by storing the training rows as context and making predictions in a single forward pass—no weight updates during fit. It accepts NumPy arrays or pandas DataFrames directly, learns embeddings for each feature, treats missing values with a dedicated vector, and groups features in sets of three to keep computation linear. The architecture splits into refinement, compression and in‑context learning stages, using cross‑attention rather than full self‑attention, which keeps GPU memory usage low. By default it runs an eight‑member ensemble with different normalizations and permutations, giving robust performance out of the box.
On TabArena Causilo tops the Elo leaderboard for single models in both classification (1771.8) and regression (2032.6), beating strong competitors like TabFM and EXAONE Tabular. ScoringBench confirms it leads on CRPS, RMSE and R² across 101 datasets. Speed tests on an H100 show Causilo fitting at 2.5 s per 1k rows and predicting at 0.25 s per 1k rows, using ~18 GiB of GPU memory—faster than TabICLv2 and TabPFN‑3 while staying within typical resource limits.
Getting started is straightforward: install via pip, import the scikit‑learn‑style classifiers or regressors, fit on your data, and predict probabilities, means, or quantiles. The first fit pulls the pretrained weights from Hugging Face automatically. For research and evaluation you can run Causilo on CUDA or CPU today; commercial deployment or a hosted API requires a separate license from Nums AI, which you can arrange after testing.
In short, Causilo removes the preprocessing bottleneck, delivers state‑of‑the‑art tabular performance with minimal code, and scales to production‑grade hardware—letting teams focus on solving business problems rather than model engineering. #AI #MachineLearning #DataScience #TabularML #OpenSource #DeployAI