unidq 0.1.0

Unified Transformer for Multi-Task Data Quality

UNIDQ: Unified Data Quality

A unified transformer architecture for multi-task data quality assessment.

UNIDQ addresses 6 data quality tasks with a single model:

• ✅ Error Detection (F1=0.894, +42% vs Raha)
• ✅ Data Repair
• ✅ Missing Value Imputation (R²=0.941, +295% vs MICE)
• ✅ Label Noise Detection (F1=0.856, +28% vs Cleanlab)
• ✅ Label Classification
• ✅ Data Valuation

Installation

pip install unidq

Quick Start

from unidq import UNIDQ, MultiTaskDataset, UNIDQTrainer

# Load your data
dataset = MultiTaskDataset(
    dirty_features=X_dirty,
    clean_features=X_clean,
    error_mask=errors,
    labels=y
)

# Initialize model
model = UNIDQ(n_features=X_dirty.shape[1])

# Train
trainer = UNIDQTrainer(model)
trainer.fit(dataset)

# Predict
results = model.predict(X_new)
print(f"Detected errors: {results['errors']}")
print(f"Imputed values: {results['imputed']}")

Citation

If you use UNIDQ in your research, please cite:

@inproceedings{unidq2026,
  title={UNIDQ: A Unified Transformer Architecture for Multi-Task Data Quality},
  author={Your Name},
  booktitle={Proceedings of the VLDB Endowment},
  year={2026}
}

License

MIT License