---
license: other
language:
- en
task_categories:
- text-classification
pretty_name: CIS 5190 News Headlines
size_categories:
- 1M<n<10M
tags:
- news
- headlines
- binary-classification
---

# CIS 5190 News Headlines

This private dataset contains cleaned news headlines collected for a binary
news-source classification project. Each row contains a normalized headline,
source label, integer label, source URL when available, date when available,
and source file provenance.

## Files

- `data/full.parquet`: canonical cleaned dataset.
- `data/balanced.parquet`: class-balanced subset.
- `data/train.parquet`, `data/validation.parquet`, `data/test.parquet`: temporal
  80/10/10 split of dated rows from the full dataset.
- `metadata/summary.json`: row counts, label mapping, date ranges, and source counts.
- `metadata/raw_manifest.json`: local raw/source file manifest used to build this export.

## Schema

- `headline`: cleaned lowercase headline text.
- `source`: source label, either `NBC` or `FoxNews`.
- `label`: integer class label, where `0 = NBC` and `1 = FoxNews`.
- `url`: source URL when available.
- `date`: publication date normalized to `YYYY-MM-DD` when available.
- `origin_file`: local source CSV that contributed the row when available.

## Summary

- Full rows: 1,499,660
- Balanced rows: 454,792
- Full source counts: `{"FoxNews": 1149644, "NBC": 350016}`
- Balanced source counts: `{"FoxNews": 227396, "NBC": 227396}`
- Full date range: `2006-01-01` to `2026-04-28`
- Rows missing URL in full data: 102
- Rows missing date in full data: 102

## Splits

The provided train/validation/test files use a temporal split over rows with
parseable dates:

- Train: 1,199,646 rows, 2006-01-01 to 2022-07-20
- Validation: 149,956 rows, 2022-07-20 to 2024-03-01
- Test: 149,956 rows, 2024-03-01 to 2026-04-28

Rows missing dates are retained in `data/full.parquet` and excluded from the
temporal split files.

## Intended Use

This dataset is intended for educational and research experiments in headline
classification, temporal generalization, domain shift, and dataset bias.

## Limitations

The dataset contains headlines and URLs only, not full article text. The source
distribution is imbalanced in the full dataset, collection coverage varies over
time, and headlines may reflect publisher-specific editorial patterns.

## Provenance

The cleaned canonical CSVs were produced by the local project pipeline in
`build_dataset.py`, which filters obvious scraping artifacts, non-article URLs,
domain/source mismatches, short headlines, and duplicate URLs/headlines.