Notnews classifiers

This repository stores the fitted classifiers used by the notnews Python package. The models classify United States and United Kingdom news text, including binary soft-news scores and detailed United States news categories.

Runtime files

The package reads four typed Parquet tables:

• metadata.parquet: model names, n-gram ranges, and class labels
• vocabulary.parquet: ordered text features
• estimators.parquet: linear coefficients and intercepts
• calibrators.parquet: isotonic calibration thresholds

The export preserves the original float64 parameters. Its validation compares feature matrices, probabilities, and labels against all three source estimators.

Source artifacts

The original scikit-learn joblib files are retained for provenance:

• us_model/nyt_us_soft_news_classifier.joblib
• us_model/nyt_us_soft_news_vectorizer.joblib
• us_model/nyt_us_classifier.joblib
• us_model/nyt_us_vectorizer.joblib
• uk_model/url_uk_classifier.joblib
• uk_model/url_uk_vectorizer.joblib

Joblib loading can execute code and is tied to the scikit-learn version used to serialize the estimators. The package does not load these files at runtime.

Training data

The package documentation links the United States model to the New York Times data work in https://github.com/notnews/us_not_news and the United Kingdom model to https://github.com/notnews/uk_not_news. This repository does not contain a complete training manifest or the underlying copyrighted article text.

Intended use

Use the models for research and exploratory classification of news content. They can support descriptive analysis of topic mix and soft-news prevalence.

Limitations

News language, section taxonomies, and publisher conventions change over time. The models may not transfer to publishers, countries, languages, or periods outside their training data. Output probabilities are model scores and should not be treated as calibrated measures of journalistic quality, importance, or truth. Do not use these models as the sole basis for moderation, employment, eligibility, or other consequential decisions.

Reproducible loading

The notnews package pins this repository to a full commit SHA and downloads the Parquet tables through huggingface_hub. Downstream code should also use the full commit SHA as revision.