Datasets:
metadata
license: cc0-1.0
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- encyclopedia
- question-answering
- text-completion
- factual
pretty_name: Encyclopedic Completion Corpus
Encyclopedic completion corpus
Year-conditioned multi-word completion pairs covering science constants, people, geography, event dates, and arithmetic. Phrases are intentionally longer than a single entity token (typically 8–20 words).
Schema
{"prompt": "The atomic number of carbon is", "phrase": "6, placing carbon among the light nonmetals…", "year": 2013, "category": "science_constants", "source": "…"}
year= knowledge cutoff for the shard (fact is valid at that cutoff)- 1000000 rows per file in
years/Y.jsonlfor Y=2013..2024
Category mix (approx.)
| category | share |
|---|---|
| science_constants | 30% |
| event_dates | 26% |
| people_founders | 22% |
| geography | 13% |
| math_arithmetic | 9% |
Sources
- Public encyclopedia-style timelines / biographies / geography
- Offline CC0-style static tables / formulas
- Synthetic arithmetic
- Optional Wikidata SPARQL (CC0)
See ANALYSIS.md for category design notes.