headlines_clean_2022_2025_slugs

A lightweight Fox News vs. NBC News binary text classification dataset built from news URLs.

Instead of using full article text, each example uses a pseudo-headline derived from the URL’s last meaningful path component (slug) and normalized into short, clean text. This makes the dataset convenient for fast baselines (TF-IDF + linear models) and also stable for transformer fine-tuning.

• Task: binary text classification (predict publisher/source)
• Labels: 0 = NBC, 1 = Fox
• Text: normalized URL slug / pseudo-headline
• Language: English
• License: MIT

---

Repository contents

This dataset repository includes two “levels” of artifacts:

1) Final modeling splits (recommended)

These three CSV files form the finalized modeling dataset with a fixed train/validation/test split:

• url_headlines_train.csv
• url_headlines_val.csv
• url_headlines_test.csv

CSV columns

Column	Type	Description
headline	string	Normalized pseudo-headline derived from the URL slug
label	int	Source label: 0 = NBC, 1 = Fox

The three splits are mutually exclusive (no duplicated headline across splits).

2) Processing / scrape log (richer intermediate table)

• headlines_clean_2022_2025_with_slug.xlsx

This XLSX file is a richer processing table (useful for auditing and reproducibility) that includes original URLs, cleaning outputs, and scrape metadata.

---

Dataset statistics

Final modeling dataset (CSV splits)

Split	#Samples	#NBC (0)	#Fox (1)	Fox (%)
Train	72,972	35,346	37,626	51.56
Val	9,121	4,418	4,703	51.56
Test	9,123	4,419	4,704	51.56
Total	91,216	44,183	47,033	51.56

The dataset is approximately balanced in each split, making accuracy a reasonable primary metric.

XLSX processing table

headlines_clean_2022_2025_with_slug.xlsx contains 91,581 rows:

• NBC: 44,374
• Fox: 47,207

Note: the XLSX table is an upstream log / intermediate artifact. The final modeling CSV splits are produced after additional normalization, filtering, and de-duplication, so counts differ slightly.

---

URL → pseudo-headline normalization

Each URL is converted into a pseudo-headline using deterministic parsing:

1. Strip query parameters and fragments (?…, #…)
2. Remove trailing .print suffixes and common file extensions when present
3. Extract the final meaningful path component as the slug
4. Replace hyphens (-) with spaces
5. Remove outlet-specific identifiers/noise (e.g., NBC patterns like rcna123456 or long trailing numeric codes)
6. Collapse extra whitespace

This produces short, clean text suitable for both sparse-vector baselines and transformer encoders.

---

XLSX schema: headlines_clean_2022_2025_with_slug.xlsx

This file contains 12 columns:

Column	Type (typical)	Description
url	string	Original news URL.
id	int	Row identifier used during collection/processing.
source	string	Outlet name (Fox or NBC).
headline	string	Headline text stored in the table (raw form as recorded by the pipeline).
scrape_status	string	Scraping status for the URL (in this release it is consistently success).
error_message	string / empty	Error message if scraping failed (empty in this release).
scrape_time	datetime	Timestamp when the URL was processed/scraped.
status_code	int	HTTP status code returned during scraping (in this release it is consistently 200).
headline_clean	string	Normalized headline text (basic cleanup for analysis/processing).
char_len	int	Character length of headline_clean.
label	int	Numeric source label used for modeling: 0 = NBC, 1 = Fox.
headline_slug	string	Slug-style version used as the modeling text input (more aggressively sanitized).

---

Quickstart

Option A: Load from the Hub with 🤗 Datasets

from datasets import load_dataset

data_files = {
    "train": "url_headlines_train.csv",
    "validation": "url_headlines_val.csv",
    "test": "url_headlines_test.csv",
}

ds = load_dataset("Evan823/headlines_clean_2022_2025_slugs", data_files=data_files)

print(ds)
print(ds["train"][0])