CoolFace
Datasetpublic

latkes/factprobe-replication-stage2-counts-canonical-v1

factprobe-replication-stage2-counts-canonical-v1 Occurrences of every probed entity name in the OLMo-2-7B mid-training corpus (576 token files), counted as OLMo token sequences with a word boundary required at each end, both written forms kept separate. SUPERSEDES factprobe-replication-stage2-counts-olmotok-v1, which was missing each entity's canonical name: the released triples name entities by their Wikidata aliases, and Wikidata keeps the canonical name in a separate field. 9… See the full description on the dataset page: https://huggingface.co/datasets/latkes/factprobe-replication-stage2-counts-canonical-v1.

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes8downloads
Dataset Card

factprobe-replication-stage2-counts-canonical-v1

Occurrences of every probed entity name in the OLMo-2-7B mid-training corpus (576 token files), counted as OLMo token sequences with a word boundary required at each end, both written forms kept separate. SUPERSEDES factprobe-replication-stage2-counts-olmotok-v1, which was missing each entity's canonical name: the released triples name entities by their Wikidata aliases, and Wikidata keeps the canonical name in a separate field. 9,183 names are new here. Verified additive: of the names counted under both name lists, not one count changed.

Dataset Info

  • —Rows: 127908
  • —Columns: 6

Columns

ColumnTypeDescription
aliasValue('string')the name searched for, exactly as written
countleadspaceValue('int64')occurrences of the space-preceded form (' Barack Obama'), the form used mid-sentence
count_bareValue('int64')occurrences of the bare form ('Barack Obama'), used after a line break, quote, bracket or hyphen
countValue('int64')the two summed. The forms partition occurrences: no single occurrence matches both, so this is addition, not double counting
corpusValue('string')which body of training data was counted
constructValue('string')the rule that produced the count, recorded so a file cannot be mistaken for one made the old way

Generation Parameters

json
{
  "experiment_name": "factprobe-replication",
  "cluster": "spark",
  "artifact_status": "final",
  "canary": false,
  "script_name": "count_npy_direct.py",
  "model": "n/a (corpus counting)",
  "description": "Occurrences of every probed entity name in the OLMo-2-7B mid-training corpus (576 token files), counted as OLMo token sequences with a word boundary required at each end, both written forms kept separate. SUPERSEDES factprobe-replication-stage2-counts-olmotok-v1, which was missing each entity's canonical name: the released triples name entities by their Wikidata aliases, and Wikidata keeps the canonical name in a separate field. 9,183 names are new here. Verified additive: of the names counted under both name lists, not one count changed.",
  "hyperparameters": {
    "names_searched": 174729,
    "names_found": 127908
  },
  "input_datasets": [
    "allenai/dolmino-mix-1124",
    "Wikidata labels, fetched 2026-08-24"
  ]
}

Usage

python
from datasets import load_dataset

dataset = load_dataset("latkes/factprobe-replication-stage2-counts-canonical-v1", split="train")
print(f"Loaded {len(dataset)} rows")