label-lens 0.1.0

Training Data Quality Analyzer — analyze labeled text classification datasets for quality issues

---

title: LabelLens emoji: 🔍 colorFrom: blue colorTo: purple sdk: docker app_file: app.py pinned: false license: mit

Label Lens

Training data quality analyzer for text classification datasets. Upload a CSV with text and label columns, get an automated quality report with actionable recommendations.

Features

• Auto-detect columns — Automatically identifies text and label columns in your CSV
• Class distribution analysis — Imbalance ratio, effective number of classes, long-tail detection, suggested focal loss weights
• Duplicate detection — Exact duplicates and near-duplicates via TF-IDF cosine similarity, with cross-class conflicts flagged as critical
• Label noise scoring — Cross-validated confidence scoring to surface likely mislabels
• Actionable report — Severity ratings (Critical/Warning/Info) with specific recommendations
• Interactive visualizations — Plotly charts for exploring your data

Quick Start

As a web app

pip install label-lens[app]
streamlit run app.py

Or with uv:

uv sync
uv run streamlit run app.py

A sample dataset is included for demo purposes.

As a library

pip install label-lens

import pandas as pd
from label_lens import (
    analyze_distribution,
    find_exact_duplicates,
    find_near_duplicates,
    score_label_noise,
    generate_report,
)

df = pd.read_csv("your_dataset.csv")  # must have 'text' and 'label' columns

dist = analyze_distribution(df)
dups = find_exact_duplicates(df)
near_dups = find_near_duplicates(df)
noise = score_label_noise(df)

report = generate_report(dist, dups, near_dups, noise)
print(report["overall_severity"])  # "Critical", "Warning", or "Info"
print(report["recommendations"])

If your CSV uses different column names, use prepare_dataframe to standardize them:

from label_lens import prepare_dataframe

df = prepare_dataframe(raw_df, text_col="content", label_col="category")

Installation

Requires Python 3.13+.

# Library only (pandas, numpy, scikit-learn)
pip install label-lens

# With Streamlit app and Plotly charts
pip install label-lens[app]

# Development
pip install label-lens[dev]

How It Works

Distribution analysis computes imbalance ratio, entropy-based effective class count, and identifies long-tail classes (<1% representation). It also calculates inverse-frequency focal loss alpha values.

Duplicate detection finds exact text matches and uses TF-IDF vectorization with chunked cosine similarity to find near-duplicates. Cross-class duplicates (same text, different labels) are flagged as critical since they represent definite labeling errors.

Noise scoring trains a logistic regression on TF-IDF features using stratified k-fold cross-validation. For each sample, it records the model's confidence in the given label. The bottom 5th percentile by confidence are flagged as mislabel suspects.

Project Structure

label_lens/
├── ingest.py         # Column detection, validation, DataFrame prep
├── distribution.py   # Class distribution analysis + visualization
├── duplicates.py     # Exact and near-duplicate detection
├── noise.py          # Label noise scoring via cross-validated confidence
├── report.py         # Aggregate findings and generate recommendations
└── utils.py          # Shared helpers

Development

# Install dev dependencies
uv sync --all-extras

# Run tests
pytest tests/ -v

# Lint and format
ruff check .
ruff format .

Deployment

Label Lens is designed for deployment on Hugging Face Spaces using Docker. The Dockerfile at the repo root handles the build.

Tech Stack

• Python 3.13+
• Streamlit
• pandas / numpy
• scikit-learn (TF-IDF, logistic regression, cross-validation)
• Plotly

License

MIT

---

Built by Mike Noe