Causilo is Nums AI’s foundation model for tabular prediction. Give it examples with known outcomes and rows whose outcomes you want to predict. The pretrained model uses those examples as context for classification or regression, without updating its weights for each new task.
The idea is to make a table the starting point for prediction: the columns describe the problem, and the known outcomes show what matters.
From context to query
Context is the table of examples whose target values are known. Query is the set of new rows with the same feature columns and unknown targets. Causilo conditions its predictions on that context.
For example, a churn task could use customer records paired with known churn outcomes as context. New customer records become queries, and the output is a predicted class or class probabilities. A task with a numeric target, such as an observed quantity, uses the regression model instead.
This is in-context learning. The model’s pretrained parameters stay fixed while the supplied examples define the prediction problem. The technical report explains this distinction and the architecture behind it.
Refine first, compress later
Causilo was pretrained on approximately 36 million synthetic tables. Its architecture lets cell representations exchange information across context rows and within each row before compressing them into row representations for prediction.
The additional refinement stage lets a cell revisit the context after incorporating information from other features. Both row refinement and compression communicate through a fixed set of summary tokens, keeping their attention cost linear in the feature count. That describes feature-side computation; total work also depends on the number of context and query rows. See the report’s architecture and scaling analysis.
Start with Python
The package exposes scikit-learn-style estimators. Install it with:
pip install causilo
For classification, supply feature rows in X_context, their observed labels in y_context, and new feature rows in X_query:
from causilo import CausiloClassifier
model = CausiloClassifier()
model.fit(X_context, y_context)
labels = model.predict(X_query)
probabilities = model.predict_proba(X_query)
Here, fit prepares the context and fitted preprocessing. It does not train new model weights. The first call also downloads and caches the checkpoint. For numeric targets, use CausiloRegressor with the same fit and predict pattern. Regression supports mean predictions and optional quantile outputs.
Inputs can be NumPy arrays or pandas DataFrames. The package supports categorical features and missing feature values, while reusing the fitted schema for query rows. Follow the repository’s runnable examples for the current environment requirements and input conventions.
Evaluate on your own data
Pretraining does not remove the need for a sound evaluation. Keep evaluation outcomes out of the context, choose splits that reflect how predictions will be used, and compare against a baseline for the same task.
Context targets must be present, even when feature values are missing. Memory and latency depend on the table, ensemble settings and device; larger inputs still need sufficient device memory. Optional attention caching trades additional memory for faster repeated predictions. The inference guide documents these boundaries.
The report evaluates accuracy and inference efficiency across TabArena, BeyondArena and ScoringBench. The repository also publishes benchmark protocols and task-level results. Those comparisons describe their stated configurations and datasets, rather than a guarantee for every table.
Code, weights and next steps
The Python code is available under Apache-2.0. The pretrained weights have a separate Causilo License v1.0. Its terms permit non-commercial research subject to the license; commercial or production use and hosted services require separate licenses. Contact contact@nums.world to discuss those uses.
Start with the code and a held-out part of your own table. Our use cases connect prediction tasks with benchmark evidence, and the technical report provides the detailed model description.