teagrjohnson/narrative-gold-annotations

NarraDolma Gold — Human Annotations

The human-annotated gold standard behind NarraBert and NarraDolma. It contains 400 passages sampled from Dolma and annotated across all 11 narrative dimensions plus event relations. This is the validation backbone of the project: every automated label (LLM and NarraBert) is measured against it.

• Paper: arXiv:2606.19468
• Collection: Narratives in LLM Pretraining Data

What's in the dataset

Each row is a 3-sentence passage with its Dolma source and human labels.

Group	Fields	Type
Agency	focalization, emotion, cognition, change_of_state, conflict	1–5 Likert
Setting	concreteness, temporal_grounding, spatial_grounding, sensory	1–5 Likert
Event relations	temporal_order, causal_relation	categorical (per event pair)

• Temporal order labels: span1_first, span2_first, simultaneous, same_event, too_hard_to_tell.
• Causal relation labels: direct_cause, enables, not_related (collapsed to a binary causal / not-causal label for downstream analysis).
• Event-relation labels are annotated at the level of adjacent event-trigger pairs, not whole passages, so a passage may contribute multiple event-pair rows. Provenance fields: dolma_id, source, topic (Common Crawl only), and a split indicator (see below).

Splits

For agency and setting, the 400 passages are divided into two non-overlapping gold sets of 200:

• Split A — used to validate the LLM annotators.
• Split B — held out to evaluate NarraBert. Event-relation annotations were too sparse to split, so event evaluation uses the full annotated set at both stages.

Annotation & agreement

One author annotated all 400 passages across the three tasks. For verification, a second author annotated overlapping subsets (agency N=100, setting N=30, events N=251) and an additional annotator participated in the setting task (N=70).

Inter-annotator agreement (per-dimension breakdowns in the paper appendix):

Task	Agreement
Agency	mean α = 0.76 (0.69–0.80), MAE 0.62
Setting	mean α = 0.70 (0.63–0.75), MAE 0.55
Event relations	mean κ = 0.68, mean F1 = 0.91

Temporal ordering is the hardest dimension (κ = 0.60), in part because of severe class imbalance among temporal relations.

Intended use & caveats

• Designed as a validation/evaluation set, not a training set — it is small by design, reflecting the fine-grained, multi-month calibration effort behind it.
• Primary annotation used for model validation was conducted by a single author; agreement on some dimensions is modest.

License & ethical considerations

Released under ODC-By. Passages are drawn from web-scraped Dolma and may include toxic, explicit, or personal content. Each row carries the Dolma unique ID for rehydration. Please use for research and auditing only.

Citation

@misc{johnson2026narrative,
  title         = {Characterizing Narrative Content in Web-scale LLM Pretraining Data},
  author        = {Johnson, Teagan and Ash, Elliott and Piper, Andrew and Antoniak, Maria},
  year          = {2026},
  eprint        = {2606.19468},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2606.19468}
}