factprobe-replication-excluded-names-v1

The complete list of probed names whose occurrence counts measure an ordinary English word rather than the entity. 1,154 names: 119 that came with He et al.'s released triples and 1035 that were added on 2026-08-24 when a quarter of the entities got their canonical Wikidata name back. Q254046 is a commune in the Ardennes called 'This'; Q1022407 and Q1809719 are both called 'Police'. Wikidata also lists song and film titles as alternative names for people, so 'She' is a name for Keyshia Cole and 'Air' for Michael Jordan. NOTHING IS EXCLUDED ANYWHERE on the strength of this file. It exists so the decision can be made when aggregating and changed without counting 15 TB again. 20 rows are marked as ones where the test itself is unreliable. SUPERSEDES the partial is_ordinary_word flag in factprobe-replication-canonical-names-added-v1, which covered only the added names.

Dataset Info

• Rows: 1154
• Columns: 8

Columns

Column	Type	Description
name	Value('string')	the probed name
came_from	Value('string')	whether it arrived with He et al.'s data or was added when the canonical names were restored
entities_named_by_it	List(Value('string'))	every Wikidata entity this is a name for. More than one means the name is ambiguous between entities as well as with the word
occurrences_as_written	Value('int64')	times it occurs in the 7B mid-training corpus exactly as written
occurrences_lower_case	Value('int64')	times the same string occurs entirely in lower case. Null when the name is itself lower case, so there is nothing to compare it against
lower_case_share	Value('float64')	lower case as a fraction of the two. Genuine names sit near zero (Alex 0.032, Washington 0.068); ordinary words near one (This 0.746, His 0.930). Flagged at 0.10
test_is_reliable_here	Value('bool')	false for short all-capitals forms, where the comparison does not mean what it means elsewhere
caution	Value('string')	why, when the test is unreliable

Generation Parameters

{
  "script_name": "build_excluded_names.py",
  "model": "n/a",
  "description": "The complete list of probed names whose occurrence counts measure an ordinary English word rather than the entity. 1,154 names: 119 that came with He et al.'s released triples and 1035 that were added on 2026-08-24 when a quarter of the entities got their canonical Wikidata name back. Q254046 is a commune in the Ardennes called 'This'; Q1022407 and Q1809719 are both called 'Police'. Wikidata also lists song and film titles as alternative names for people, so 'She' is a name for Keyshia Cole and 'Air' for Michael Jordan. NOTHING IS EXCLUDED ANYWHERE on the strength of this file. It exists so the decision can be made when aggregating and changed without counting 15 TB again. 20 rows are marked as ones where the test itself is unreliable. SUPERSEDES the partial is_ordinary_word flag in factprobe-replication-canonical-names-added-v1, which covered only the added names.",
  "hyperparameters": {
    "threshold_lower_case_share": 0.1,
    "minimum_occurrences_to_test": 1000,
    "names_flagged": 1154
  },
  "input_datasets": [
    "allenai/dolmino-mix-1124",
    "Zenodo 10.5281/zenodo.15092789"
  ],
  "experiment_name": "factprobe-replication",
  "cluster": "spark",
  "artifact_status": "final",
  "canary": false
}

Usage

from datasets import load_dataset

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

---