Orion-BiX is an advanced tabular foundation model that combines Bi-Axial Attention with Meta-Learning capabilities for few-shot tabular classification. The model extends the TabICL architecture with alternating attention patterns and episode-based training, achieving state-of-the-art performance on domain-specific benchamrks such as Healthcare and Finance.

Orion-BiX introduces three key architectural innovations:

1. Bi-Axial Attention: Alternating attention patterns (Standard → Grouped → Hierarchical → Relational) that capture multi-scale feature interactions
2. Meta-Learning: Episode-based training with k-NN support selection for few-shot learning
3. Configurable Architecture: Flexible design supporting various attention mechanisms and training modes
4. Production Ready: Memory optimization, distributed training support, and scikit-learn interface

Orion-BiX follows a three-component architecture:

Input → Column Embedder (Set Transformer) → Bi-Axial Attention → ICL Predictor → Output

1. Column Embedder: Set Transformer for statistical distribution learning across features from TabICL
2. Bi-Axial Attention: Replaces standard RowInteraction with alternating attention patterns:
   • Standard Cross-Feature Attention: Direct attention between features
   • Grouped Feature Attention: Attention within feature groups
   • Hierarchical Feature Attention: Hierarchical feature patterns
   • Relational Feature Attention: Full feature-to-feature attention
   • CLS Token Aggregation: Multiple CLS tokens (default: 4) for feature summarization
3. tf_icl ICL Predictor: In-context learning module for few-shot prediction

Each BiAxialAttentionBlock applies four attention patterns in sequence:

Standard → Grouped → Hierarchical → Relational → CLS Aggregation

• Python 3.9-3.12
• PyTorch 2.2+ (with CUDA support recommended)
• CUDA-capable GPU (recommended for training)

cd orion-bix
pip install -e .

pip install git+https://github.com/Lexsi-Labs/Orion-BiX.git

Orion-BiX provides a scikit-learn compatible interface for easy integration:

from orion_bix.sklearn import OrionBixClassifier

# Initialize and fit the classifier
clf = OrionBixClassifier()

# Fit the model (prepares data transformations)
clf.fit(X_train, y_train)

# Make predictions
predictions = clf.predict(X_test)
probabilities = clf.predict_proba(X_test)

Orion-BiX includes automatic preprocessing that handles:

1. Categorical Encoding: Automatically encodes categorical features using ordinal encoding
2. Missing Value Imputation: Handles missing values using median imputation for numerical features
3. Feature Normalization: Supports multiple normalization methods:
   • "none": No normalization
   • "power": Yeo-Johnson power transform
   • "quantile": Quantile transformation to normal distribution
   • "quantile_rtdl": RTDL-style quantile transform
   • "robust": Robust scaling using median and quantiles
4. Outlier Handling: Clips outliers beyond a specified Z-score threshold (default: 4.0)
5. Feature Permutation: Applies systematic feature shuffling for ensemble diversity:
   • "none": Original feature order
   • "shift": Circular shifting
   • "random": Random permutation
   • "latin": Latin square patterns (recommended)

The preprocessing is automatically applied during fit() and predict(), so no manual preprocessing is required.

If you use Orion-BiX in your research, please cite our paper:

@article{bouadi2025orionbix,
      title={Orion-Bix: Bi-Axial Attention for Tabular In-Context Learning}, 
      author={Mohamed Bouadi and Pratinav Seth and Aditya Tanna and Vinay Kumar Sankarapu},
      year={2025},
      eprint={2512.00181},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2512.00181}, 
}

This project is released under the MIT License. See LICENSE for details.

For questions, issues, or contributions, please:

• Open an issue on GitHub
• Join our Discord community

Orion-BiX is built on top of TabICL, a tabular foundation model for in-context learning. We gratefully acknowledge the TabICL authors for their foundational work and for making their codebase publicly available.