CoolFace
Datasetpublic

latkes/factprobe-replication-canonical-names-added-v1

factprobe-replication-canonical-names-added-v1 The canonical name that was missing for 10,471 of the 42,235 probed entities (24.8%), and had therefore never been counted in the corpus nor put to the model. He et al.'s released triples name entities by their Wikidata aliases -- the 'also known as' field -- which by construction excludes the canonical name. Q76 carried 31 names, none of them 'Barack Obama'. Each row also records whether the added name is an ordinary English word… See the full description on the dataset page: https://huggingface.co/datasets/latkes/factprobe-replication-canonical-names-added-v1.

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes4downloads
Dataset Card

factprobe-replication-canonical-names-added-v1

The canonical name that was missing for 10,471 of the 42,235 probed entities (24.8%), and had therefore never been counted in the corpus nor put to the model. He et al.'s released triples name entities by their Wikidata aliases -- the 'also known as' field -- which by construction excludes the canonical name. Q76 carried 31 names, none of them 'Barack Obama'. Each row also records whether the added name is an ordinary English word, tested by how often it is written in lower case in the same corpus.

Dataset Info

  • —Rows: 10471
  • —Columns: 7

Columns

ColumnTypeDescription
entityValue('string')Wikidata item, e.g. Q76
canonical_nameValue('string')the name added, from Wikidata's label field
occurrences_stage2Value('int64')how often it occurs in the 7B mid-training corpus as written; null if under 1,000 and so not tested
occurrenceslowercaseValue('int64')how often the same string occurs entirely in lower case; null when the name is already lower case and nothing can be compared
lowercaseshareValue('float64')lower case as a fraction of the two. Real names sit near zero (Alex 0.032, Washington 0.068); ordinary words sit near one (This 0.746, His 0.930)
isordinarywordValue('bool')true above a 0.10 share, or when the name is itself written in lower case. Nothing is excluded anywhere on this basis -- it is recorded so the decision can be made and changed without counting again
whyValue('string')which of those two grounds applied

Generation Parameters

json
{
  "experiment_name": "factprobe-replication",
  "cluster": "spark",
  "artifact_status": "final",
  "canary": false,
  "script_name": "add_canonical_names.py",
  "model": "n/a",
  "description": "The canonical name that was missing for 10,471 of the 42,235 probed entities (24.8%), and had therefore never been counted in the corpus nor put to the model. He et al.'s released triples name entities by their Wikidata aliases -- the 'also known as' field -- which by construction excludes the canonical name. Q76 carried 31 names, none of them 'Barack Obama'. Each row also records whether the added name is an ordinary English word, tested by how often it is written in lower case in the same corpus.",
  "hyperparameters": {
    "entities": 42235,
    "entities_repaired": 10471,
    "distinct_names_added": 10186
  },
  "input_datasets": [
    "Zenodo 10.5281/zenodo.15092789",
    "Wikidata labels, fetched 2026-08-24"
  ]
}

Usage

python
from datasets import load_dataset

dataset = load_dataset("latkes/factprobe-replication-canonical-names-added-v1", split="train")
print(f"Loaded {len(dataset)} rows")