NPOV-Onset

A dated benchmark of neutral-point-of-view (NPOV) dispute onsets on English Wikipedia, with topically and activity-matched never-tagged controls.

For each of ~2,000 articles that received an NPOV maintenance template, the onset is localised by binary search to the exact revision that introduced the tag — correcting for the ~26% of templates that are only bot-dated and for edit-summary regexes that recover barely half of human tag additions. Each event is paired with a never-tagged article that shares a Wikipedia category and has comparable edit activity, so the comparison is fair.

• Code & full documentation: https://github.com/mahi01agarwal/npov-drift-detector
• Datasheet (Gebru et al. format): see DATASHEET.md in the code repository
• Tracking task: phab:T428896
• License: CC BY-SA 4.0 (derived from English Wikipedia)

Configurations

Config	Rows	What it is
events	2,013	NPOV-dispute onsets, each pinned to its onset revision
controls	688	Matched never-tagged articles (688 matched pairs; join on matched_to)
localisation_attempts	2,166	The full upstream localisation log (≈93% located), for auditing the ground truth

Usage

from datasets import load_dataset

events   = load_dataset("MahimaSinghal/npov-onset", "events",   split="train")
controls = load_dataset("MahimaSinghal/npov-onset", "controls", split="train")

# Reconstruct matched pairs (one control per event, join on title):
ctrl_by_event = {c["matched_to"]: c for c in controls}
pairs = [(e, ctrl_by_event[e["title"]]) for e in events if e["title"] in ctrl_by_event]
print(len(pairs), "matched pairs")   # 688

The dataset ships metadata and revision ids, not article text. To get the prose for any snapshot, fetch the onset_revid (or an earlier revision for lead-time work) from the MediaWiki Action API; the repository's scripts do this and cache the results.

Key fields

events — title, onset_revid/onset_ts (the localised onset), onset_user, onset_comment, template (pov, npov, pov section, …), scope (article/section), anchor_is_bot_dating (flag: anchor was a bot dating the template), first_tag_censored (flag: tag predates the observation window), article_chars. controls — title, pageid, length, matched_to (the event it controls for), match_category, match_distance.

See the repository DATASHEET.md for the complete data dictionary, collection process, recommended splits (by matched pair and/or temporally, to avoid leakage), and intended-use limits.

Intended use and limits

Built for concurrent detection and lead-time / drift analysis of NPOV disputes, and for studying measurement bias in Wikipedia maintenance templates. The label is "received an NPOV template" — a proxy for an editor's judgment, not a gold neutrality label. English only; the 688 controls are a matched subset, not a random sample.

Do not use this to profile or draw conclusions about individual editors named in onset_user/edit summaries (these are already public in page history), or to drive automated edits on live articles.

Citation

@misc{npov_onset_2026,
  title        = {NPOV-Onset: a dated benchmark of neutral-point-of-view dispute onsets on English Wikipedia},
  author       = {Agarwal, Mahima and Dev},
  year         = {2026},
  howpublished = {\url{https://github.com/mahi01agarwal/npov-drift-detector}},
  note         = {CC BY-SA 4.0; Wikimedia Phabricator T428896}
}