1.3 million news articles and summaries across 38 major publications. Metadata includes publication date and URL, alongside quantitative metrics like extractive density and compression ratios.
Use Cases
- Train summarization models using the text and summary fields
- Analyze the relationship between article length and summary quality using the compression and density features
- Perform longitudinal studies of news reporting using the date and title columns
- Filter training data by extractive or abstractive style using the coverage_bin feature
Strengths
- 1.3 million article-summary pairs from 38 distinct news organizations
- Quantitative metrics for summarization analysis including density, coverage, and compression
- Categorical labels in density_bin, coverage_bin, and compression_bin for filtering by low, medium, or high summary styles
- Standardized data splits provided in train.jsonl, dev.jsonl, and test.jsonl formats