wiki-model-facts / README.md
Lucasgitguy-co's picture
wiki-model 0.1.0
eb8d17d verified
|
Raw History Blame Contribute Delete
2.41 kB
metadata
license: cc-by-sa-4.0
language:
  - en
pretty_name: wiki-model facts
tags:
  - wikipedia
  - facts
  - question-answering
  - retrieval
  - bm25
size_categories:
  - 1M<n<10M
source_datasets:
  - wikimedia/structured-wikipedia
configs:
  - config_name: tasks
    data_files:
      - split: train
        path: tasks/train.jsonl.gz
      - split: validation
        path: tasks/val.jsonl.gz
      - split: test
        path: tasks/test.jsonl.gz
  - config_name: articles
    data_files: articles/wiki.jsonl.gz

wiki-model facts

Training data and the sourced fact index for wiki-model and wiki-model-tiny, built from wikimedia/structured-wikipedia (English, 2024).

Path What
articles/wiki.jsonl.gz 947,727 articles: title, url, desc, sents (up to 3 lead sentences), fields (selected infobox facts), src
tasks/{train,val,test}.jsonl.gz task (fact / q / fix), title, prompt, target, ctx (context text or null). Split by article title.
index/ bm25s index (bm25/) + docs.sqlite (article text, zlib) used by --grounded

Articles: 232k random articles from three shards, plus 716k long articles (at least 15k characters, no digits in the title) from the other 83 shards.

How tasks are built
  • fact: <fact> Title </fact> → the first one or two lead sentences.
  • q: <q> What is Title? </q> → the short description; <q> When was Title founded? </q> etc. → an infobox value (about 60 fields).
  • fix: an abstract sentence with a year shifted by 1–30 years → the original; or an unchanged sentence → itself.
  • ctx: 35% of examples carry <ctx> text: the article's own index text (same format as the index), half the time together with a random distractor article. At run time --grounded fills it from BM25 search instead.
  • The models were also trained on Bulutthecat/FactfulDataset fix examples. They are not included here, because that dataset has no licence on its card.

Licence and attribution

CC BY-SA 4.0. Text and data come from Wikipedia, written by Wikipedia contributors, via the Wikimedia Enterprise structured contents snapshot. Each article row keeps its source URL.