DescribeEvents/wikipedia_edit_events

Wikipedia Edit Events

Curated Wikipedia article edit history sequences from Featured and Good articles, designed for temporal point process (TPP) and marked temporal point process (MTPP) modeling. Each sequence tracks an article's editorial activity over time, where the prediction target is the edit action type.

Schema

Each record is a dictionary with 8 fields:

Field	Type	Description
seq_idx	int	Sequence index
seq_len	int	Number of events in the sequence
description	str	Natural language description of the article's edit timeline
metadata	str (JSON)	article_title, article_pageid, chunk_index, num_editors, date_range_start, date_range_end, time_unit
time_since_start	list[float]	Time since the first event (in hours)
time_since_last_event	list[float]	Time since the previous event (in hours)
type_event	list[str]	Edit action type label (see below)
type_text	list[str]	Natural language description of each edit

Event Types (6 edit action types)

Edits are classified by priority order from revision metadata (byte delta, tags, edit summary):

Type	Description	Classification Rule
revert	Undoing a previous edit	Tags contain mw-rollback, mw-undo, or mw-manual-revert
reference	Adding or modifying citations	Edit summary matches ref/cite/source/bibliography patterns
content_addition	Substantial new content	Byte delta > +200
content_removal	Substantial content removed	Byte delta < -200
copy_edit	Small wording/formatting changes	|delta| <= 200, not flagged as minor
minor_edit	Trivial corrections	Minor flag set

Curation Filters

Sequences are aggressively filtered to ensure diverse, non-trivial edit type distributions:

Filter	Value	Description
max-articles	1,000	Featured + Good articles processed
min-seq-size	60	Min events per sequence chunk
max-seq-size	80	Max events per sequence chunk
min-types	4	At least 4 distinct edit types per sequence
max-duration-hours	4,380	Drop sequences spanning more than 6 months
max-dominant-ratio	0.40	Drop sequences where any single type exceeds 40%
max-repeat-ratio	0.40	Drop sequences where consecutive repeat ratio exceeds 40%

Additional processing:

• Bot edits excluded (tag "bot" or username ending in "bot")
• Byte deltas computed from consecutive revision sizes
• Sequences chunked chronologically per article

Event Text

Each event's type_text is a coherent natural sentence providing rich context without explicitly stating the edit type classification. The article title is omitted (available at sequence level in description and metadata).

Registered editor My love is love made a moderate addition (+620 bytes, <1% of article) to section "Titling and artwork"

Sentence structure: {Editor identity} {size action} [to section "..."]. [Summary: "..."]. [Signals]. [Tool]. [Minor flag].

Components (included only when present):

1. Editor identity (lead) — "Registered editor Centrx" or "An anonymous editor"
2. Size action (verb) — "made a substantial addition (+1,253 bytes, 61% of article)"
3. Section target — to section "Background and formation" (extracted from edit comment)
4. Edit summary — Summary: "start background" (cleaned of wiki markup)
5. Semantic signals — auto-detected patterns (e.g., "The summary mentions fixing language or typography")
6. Tool/interface — "Made using the visual editor" (from revision tags)
7. Minor flag — "Flagged as minor"

The model must learn to predict the edit action type from these contextual signals.

Example

{
  "seq_idx": 2,
  "seq_len": 80,
  "description": "Wikipedia edit timeline for the article '4 (Beyoncé album)' spanning Dec 2011 to Jan 2012. This sequence tracks editorial activity including content changes, reverts, and maintenance edits.",
  "metadata": "{\"article_title\": \"4 (Beyoncé album)\", \"article_pageid\": 31176601, \"chunk_index\": 24, \"num_editors\": 22, \"date_range_start\": \"2011-12-13T16:21:57+00:00\", \"date_range_end\": \"2012-01-07T21:12:43+00:00\", \"time_unit\": \"hours\"}",
  "time_since_start": [0.0, 0.22, 2.56, ...],
  "time_since_last_event": [0.0, 0.22, 2.35, ...],
  "type_event": ["content_addition", "minor_edit", "copy_edit", ...],
  "type_text": [
    "Registered editor My love is love made a moderate addition (+620 bytes, <1% of article) to section \"Titling and artwork\"",
    "Registered editor My love is love made a null edit (no size change) to section \"Music and themes\". Flagged as minor",
    "Registered editor Dumb + dumb = 4 made a small removal (-142 bytes)",
    ...
  ]
}

Intended Use

• Training and evaluating temporal point process models
• Studying collaborative editing patterns on Wikipedia
• Benchmarking next-event prediction and event forecasting
• Modeling marked temporal point processes with rich text marks

Citation

If you use this dataset, please cite:

@dataset{wikipedia_edit_events_2025,
  title={Wikipedia Edit Events},
  author={XiaoBB},
  year={2025},
  url={https://huggingface.co/datasets/DescribeEvents/wikipedia_edit_events},
  note={Curated from Wikipedia MediaWiki API, Featured and Good articles}
}