Datasets:

CompassioninMachineLearning
/

paw-wikipedia-attribution

dataset_info:
  - config_name: paw_edits
    features:
      - name: text
        dtype: string
      - name: page
        dtype: string
      - name: section_name
        dtype: string
      - name: is_paw_edit
        dtype: bool
  - config_name: queries
    features:
      - name: query
        dtype: string
      - name: type
        dtype: string
license: cc-by-4.0
task_categories:
  - text-classification
tags:
  - animal-welfare
  - wikipedia
  - data-attribution
  - llm-training

PAW Wikipedia Attribution Dataset

Data for the paper "Small edits, large models: How Wikipedia advocacy shapes LLM values" (Navas, Brazilek & Gnauck, 2026).

Files

paw_edits.csv (236 rows)

Wikipedia sections used in training data attribution experiments. Each row is a text chunk from Wikipedia.

Column	Description
`text`	The Wikipedia section text
`page`	Wikipedia article title
`section_name`	Section heading
`is_paw_edit`	Whether this section was added/edited by Pro-Animal Wikipedians

118 PAW-edited sections (animal welfare content added by advocacy editors) paired with 118 control sections from the same Wikipedia articles.

queries.csv (170 rows)

Queries used to measure training data influence.

Column	Description
`query`	The query text
`type`	`animal_welfare` (80 queries) or `general` (90 queries)

Animal welfare queries ask about animal welfare at specific companies/by specific politicians. General queries ask about the same entities without mentioning animal welfare.

Citation

@article{navas2026small,
  title={Small edits, large models: How Wikipedia advocacy shapes LLM values},
  author={Navas, M and Brazilek, J and Gnauck, A},
  year={2026}
}