README.md

---
license: apache-2.0
tags:
- anomaly-detection
- tabular-data
- in-context-learning
- one-class-classification
- unsupervised-learning
- semi-supervised-learning
---

# ICLAD Pretrained Checkpoints

**ICLAD: In-Context Learning for Unified Tabular Anomaly Detection**

This repository contains pretrained model checkpoints for ICLAD, an in-context learning framework for tabular anomaly detection that supports one-class, unsupervised, and semi-supervised settings.

📄 **Paper:** [ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes](https://arxiv.org/abs/2603.19497)

## Model Variants

### 1. [ICLAD]  `iclad_mixedprior_unified.pth` (Default)

General use across all anomaly detection scenarios

- Trained on **mixed prior** (structural causal models (SCMs) and perturbation noises)
- Supports all three settings: **one-class**, **unsupervised**, and **semi-supervised**
- **Recommended as the default choice** for most use cases

---

### 2. [ICLAD_OC]  `iclad_mixedprior_oneclass.pth`

Used for ablation studies.

- Trained on **mixed prior** optimized for **one-class setting only**
- **Note:** Use `iclad_mixedprior_unified.pth` for general one-class applications

---

### 3. [ICLAD_UNSUP]  `iclad_mixedprior_unsup.pth`

Used for ablation studies.

- Trained on **mixed prior** optimized for **unsupervised setting only**
- **Note:** Use `iclad_mixedprior_unified.pth` for general unsupervised applications

---

### 4. [ICLAD_SCM]  `iclad_scm_unified.pth`

Used for ablation studies.

- Trained on **SCM-only prior** (no perturbation-based noise)
- Supports all three settings: **one-class**, **unsupervised**, and **semi-supervised**
- **Note:** Use `iclad_mixedprior_unified.pth` for general applications

---

## Usage

### ⭐ Recommended: Use the Unified Model

**For most applications, use `iclad_mixedprior_unified.pth`** (the default). The other variants are provided for research and paper reproduction purposes.

### Option 1: Load from Hugging Face Hub

```python
from iclad import ICLAD

# Load default unified model (works for all settings)
model = ICLAD.from_checkpoint("jyiwei/iclad-checkpoints/iclad_mixedprior_unified.pth")
```

### Option 2: Load from Local Checkpoints

If you have the checkpoint files in `src/iclad/checkpoints/`, use the built-in names:

```python
from iclad import ICLAD

# ⭐ RECOMMENDED: Load default unified model
model = ICLAD()  # Uses iclad_mixedprior_unified by default
model = ICLAD(model_name="ICLAD")

# --- Paper Reproduction Only (below) ---

# Load one-class variant
model = ICLAD(model_name="ICLAD_OC")

# Load unsupervised variant
model = ICLAD(model_name="ICLAD_UNSUP")

# Load SCM-only variant
model = ICLAD(model_name="ICLAD_SCM")
```

### Example: Anomaly Detection on Tabular Data

```python
import numpy as np
from iclad import ICLAD

# Initialize model
model = ICLAD(model_name="ICLAD")  # Default unified model

# Prepare training and test data
X_train = np.random.randn(100, 10)  # 100 samples, 10 features
X_test = np.random.randn(50, 10)    # Test data

# Fit on training data (for unsupervised setting, no labels needed)
model.fit(X_train)

# Get anomaly scores
scores = model.predict_score(X_test)
```

### One-class Example

```python
# Prepare training and test data
X_train = np.random.randn(100, 10)  # 100 samples, 10 features
Y_train = np.zeros(X_train.shape[0]) # All normal
X_test = np.random.randn(50, 10)    # Test data

# Fit on training data (for one-class setting, all labels should be zero)
model.fit(X_train, Y_train)

# Get anomaly scores
scores = model.predict_score(X_test)
```

### Semi-Supervised Example

```python
# y_train: 1 for anomaly, -1 for unknown (no support for known normals yet)
model.fit(X_train, Y_train)
scores = model.predict_score(X_test)
```

## Citation

If you use these pretrained models in your research, please cite:

```bibtex
@misc{wei2026icladincontextlearningunified,
  title={ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes},
  author={Jack Yi Wei and Narges Armanfard},
  year={2026},
  eprint={2603.19497},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2603.19497},
}
```

## License

The model code is licensed under the **Apache License 2.0**.

## Repository

For issues, discussions, and more information, please visit the main ICLAD repository.

## Contact

Jack Wei: yi.wei4@mail.mcgill.ca