CNN/DailyMail: 100K-1M Non-Anonymized News Summarization Pairs

Aggregating between 100,000 and 1,000,000 non-anonymized news articles and summaries sourced from CNN and Daily Mail. Curated by ccdv and last updated in 2022, it provides paired text for training and evaluating abstractive summarization models.

Use Cases

• Training sequence-to-sequence models to generate 'highlights' from 'article' text
• Evaluating ROUGE or METEOR scores for abstractive summarization algorithms
• Fine-tuning large language models on news-style prose and formatting

Strengths

• Scale of 100,000 to 1,000,000 records
• Non-anonymized text preserving original entity names and locations
• Apache 2.0 license for permissive usage

Limitations

• Geographic bias toward United States news sources
• Potential for lead bias where the beginning of the 'article' text contains the majority of summary information
• Relies on found data from website highlights rather than expert-annotated summaries

Provenance

Source

CNN and Daily Mail websites

Collection Method

Found and scraped from original news sources

Freshness

Last updated October 2022; no ongoing update frequency specified.

Geography

United States