FACTPROP

Popular Knowledge Propagates More Errors in LLM Knowledge Updating

Project page & demo · Research code

Overview

FACTPROP is a factual graph for studying how knowledge updates affect connected facts in large language models. This initial release provides the entity popularity index used by the project’s browser explorer: 100,015 graph-node records with entity labels, available Wikidata QIDs, and forward-edge object in-degree.

This is an index release, not the full factual graph. Relation triples, natural-language QA, evidence passages, and experiment outputs are not included. The full graph package is being prepared separately.

Contents

File	Description
data/entities.jsonl	One object per graph node, suitable for the dataset viewer
browser-index.json	Unchanged index used by the project website
metadata/graph.json	Graph counts, thresholds, export date, and source-graph checksum
metadata/release.json	Source version and mapping statistics
SHA256SUMS	File integrity checksums

The entities split is the complete entity index. It is not a training or test split.

Fields

Field	Type	Meaning
label	string	Original graph-node label
qid	string or null	Wikidata QID where mapped
forward_object_in_degree	integer	Number of forward factual edges pointing to this node
qid_node_count	integer	Number of index records sharing this non-null QID; zero when QID is missing

Graph-node labels, rather than QIDs, distinguish records. The added qid_node_count field describes mapping multiplicity; it does not alter the original scores.

Structural popularity

For an entity o, popularity counts incoming factual edges whose is_inverse attribute is not true. The index represents 357,205 forward edges; the original graph also stores 75,357 inverse edges, for 432,562 stored edges in total.

This score measures connectivity in this graph version. It does not measure search traffic or public familiarity, and it does not predict a model’s error rate for a particular input.

For example, Apple Inc. / Q312 has degree 467, while Apple / Q89 has degree 4.

Load the index

from datasets import load_dataset

entities = load_dataset("factprop/FACTPROP", "entity_index", split="entities")
matches = entities.filter(lambda row: row["qid"] == "Q312")
print(matches[:])

To retrieve the browser-format file:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("factprop/FACTPROP", "browser-index.json", repo_type="dataset")
with open(path, encoding="utf-8") as stream:
    index = json.load(stream)

Mapping coverage and limitations

• 33,901 graph-node records have no QID mapping.
• 5,037 QIDs correspond to more than one graph-node record.
• Some expected mappings, including the Java and Python programming languages, are absent.
• A missing mapping does not imply low popularity or a score of zero. A known graph node may independently have a true degree of zero.
• Multiple nodes with one QID must be reviewed separately. Do not automatically sum their degrees or take the maximum.
• This index supports node-level lookup. It cannot reconstruct the full graph or reproduce all experiments in the paper.

Provenance and versions

The files derive from the index already published in the FACTPROP code repository. That index was exported from the paper graph; its source checksum is retained in metadata/graph.json. The graph was constructed from Wikipedia-derived facts with Wikidata validation, as described in the paper.

Entity labels, QIDs, and scores are preserved without remapping or aggregation. Proposed mapping corrections should be maintained separately from the original records.

License

A dataset-specific license has not yet been specified by the authors. This index release does not assign a new license to its upstream sources. Licensing and attribution for the full graph package remain under review.

Citation

The public preprint and final bibliographic entry will be added when available. Please refer to the project under the paper title:

Popular Knowledge Propagates More Errors in LLM Knowledge Updating