omnilint 0.1.5

AI Dataset Quality Auditor - Tabular & Image

OmniLint

AI Dataset Quality Auditor — detect, report, and score quality issues in tabular and image datasets before they break your model.

Overview

OmniLint analyzes raw datasets (CSV/Parquet for tabular, COCO/YOLO for images) and surfaces actionable quality issues with a composite Data Quality Score (DQS). It serves as a gatekeeper between raw data and ML pipelines.

Features

Tabular Auditing

• Basic Checks — Missing values, data type mismatches, constant columns, duplicates
• Distribution Analysis — Skewness detection, IQR/Z-score outlier identification
• Label Auditing — Class imbalance detection, rare class warnings
• Leakage Detection — High correlation with target (Pearson/Spearman), categorical leakage proxies (Cramér's V)
• Feature Importance — RF probe model for noise/suspiciously powerful features
• Duplicate Detection — Exact duplicates and near-duplicate (cosine similarity/FAISS) matching

Image Auditing (Beta)

• Integrity Checks — Corrupt files, resolution outliers, format inconsistency
• Distribution Analysis — Brightness, contrast, color channel imbalance
• Label Checks — Label-file mismatch, class imbalance
• Duplicate Detection — Perceptual hash (pHash) exact duplicates, CLIP near-duplicates
• Anomaly Detection — Blur, exposure issues, blank images

Installation

Base (Tabular Only)

pip install omnilint

With Image Support

pip install omnilint[image]

Development

pip install -e ".[dev,image]"

Quick Start

CLI - Tabular

OmniLint run dataset.csv --target label --output report.html

CLI - Image

OmniLint run dataset/ --mode image --format yolo --output report.json

Python API

# Tabular
from omnilint.core import load, AuditEngine
from omnilint.core.engine import AuditConfig

df, schema = load("dataset.csv")
config = AuditConfig(target_col="label")
engine = AuditEngine(df, config)
result = engine.run()

print(result.quality_score)  # e.g., 73.4

# Image
from omnilint.core.loader import load

image_dataset = load("path/to/coco_data")
image_config = AuditConfig(mode="image")
image_engine = AuditEngine(image_dataset, image_config)
image_result = image_engine.run()

print(image_result.quality_score)  # e.g., 85.2

Streamlit UI

streamlit run app/streamlit_app.py

Input Formats

Tabular

• CSV (.csv)
• Parquet (.parquet)

Image

• COCO JSON (annotations in JSON with images/annotations/categories)
• YOLO (folder with data.yaml or train/images/train/labels structure)

Score Bands

Score	Band	Action
0–40	Critical	Do not train
41–65	Poor	Major fixes required
66–80	Fair	Minor fixes recommended
81–100	Good	Ready for training

Image Mode Weights

Module	Weight
Integrity	30%
Duplicates	25%
Anomalies	20%
Labels	15%
Distribution	10%

Project Structure

OmniLint/
├── omnilint/
│   ├── core/           # Engine, loader (CSV/COCO/YOLO), scorer
│   ├── tabular/        # Tabular audit modules
│   │   ├── checks/     # basic, distribution, labels, leakage, importance, dedup
│   │   ├── report/     # Report builders
│   │   └── utils/      # Stats, embeddings
│   └── image/          # Image audit modules (beta)
│       ├── checks/     # integrity, distribution, labels, duplicates, anomalies
│       └── utils/      # phash, clip_encoder, pixel_stats
│
├── cli/                # Typer CLI
├── app/                # Streamlit UI
├── app/components/     # Tabular UI components
├── app/components/image/ # Image UI components
└── tests/
    ├── tabular/        # Tabular tests
    └── image/          # Image tests

Tech Stack

Core

• Core: Python 3.11+, Pandas, NumPy, SciPy
• ML: scikit-learn, FAISS, sentence-transformers
• CLI: Typer, Rich
• UI: Streamlit, Plotly
• Reports: Jinja2

Image (Optional)

• I/O: Pillow, OpenCV
• Hashing: imagehash (pHash)
• Embeddings: OpenAI CLIP (ViT-B/32)
• Blur: OpenCV Laplacian

See full release history in CHANGELOG.md

License

MIT