NarraDolma: Narrative Feature Vectors for the 3-Trillion-Token Dolma Corpus

NarraDolma provides a large-scale narrative characterization of the Dolma pretraining corpus. It contains approximately 3 million passages drawn from about 785,000 unique documents across all 12 Dolma sub-corpora, each labeled with a fine-grained narrative feature vector. The dataset was created by teagrjohnson and is intended as a resource for studying how narrative qualities are distributed in web-scale data.

Use Cases

• Analyze the distribution of narrative qualities across different web domains based on the 12 Dolma sub-corpora.
• Study the relationship between narrative features and model performance on downstream tasks.
• Train or evaluate models for narrative understanding or generation based on the fine-grained feature vectors.
• Conduct comparative studies of narrative content across different data sources within the Dolma corpus.

Strengths

• Large scale with ~3 million passages labeled with narrative features.
• Broad coverage across ~785,000 unique documents from all 12 Dolma sub-corpora.
• Provides fine-grained narrative feature vectors produced by NarraBert.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Passages drawn from the Dolma corpus and labeled with narrative feature vectors by NarraBert.

Freshness

Last updated 2026-06-19 19:02:17; freshness should be verified.