README.md

metadata

license: cc-by-4.0
language:
  - en
tags:
  - knowledge-graph
  - rdf
  - n-triples
  - parquet
  - llm-generated
pretty_name: GPTKB v2
size_categories:
  - 10M<n<100M

GPTKB v2

Browse, query or ask the full knowledge base at gptkb.org.

Two artifact lines, one knowledge base (snapshot gpt51_270226, 38,450,135 source facts):

File	What it is
gptkb_v2.0.1_rdf.nt.gz	Simplified RDF export. Use this one.
gptkb_v2.0.1_full_record/	The full six-term record, as Parquet.
gptkb_v2.0_rdf.nt.gz	Superseded interim RDF export, kept for citability.

Simplified RDF export (gptkb_v2.0.1_rdf.nt.gz)

Format	N-Triples, gzipped (425 MB; 4.1 GB uncompressed)
Size	42,327,043 lines → 41,511,580 distinct triples once loaded
sha256	202d9e32f28548ae661b3f8cf9db83411fe18a063702f9c98df4b553d750e801
License	CC BY 4.0

The graph served by the SPARQL endpoint at gptkb.org/query. Entities, predicates and concepts are ID-based IRIs:

<https://gptkb.org/entity/E0>    a person, place, work, …
<https://gptkb.org/prop/P0>      a predicate
<https://gptkb.org/concept/C0>   a class

The vocabulary is deliberately three predicates and no more:

Predicate	Meaning
<https://gptkb.org/prop/P…>	a disambiguated fact
rdfs:label	the canonical name of an entity, predicate or concept
skos:altLabel	every surface form observed for an entity

Literals are plain RDF strings; typed literals (xsd:date, xsd:integer) are deferred. Concepts carry no rdfs:subClassOf: the source has no hierarchy to export.

v2.0.1 vs the interim v2.0 file. Same knowledge base, same snapshot: the difference is how much of it the pipeline had disambiguated at export time. The interim file was cut when roughly half of the named-entity objects had been resolved to entity IDs; facts whose object was still unresolved could mint no IRI and were skipped. This re-export was cut after the disambiguation pipeline completed: about one million further entities exist in the catalog, named-entity coverage is ~99.8%, and only 62,600 of the 38,450,135 source facts remain unexportable. Concretely: the loaded graph grows from 32.8M to 41.5M triples, and E14 (United States of America) now carries 168 skos:altLabel surface forms where the interim file had 127.

Loading

# Virtuoso, QLever, Jena, rdflib: anything that reads N-Triples
gunzip -c gptkb_v2.0.1_rdf.nt.gz | head

Load into a named graph of <https://gptkb.org/gptkb> to match the IRIs the file already carries.

The full record (gptkb_v2.0.1_full_record/)

A fact in GPTKB v2 has six terms, not three: alongside the disambiguated subject, predicate and object, it records the wording each was stated with, and the provenance of how that wording was resolved. A triple has nowhere to put either. In the RDF export, <E0> <P0> <C0> appears five times because five differently worded facts about Vannevar Bush resolve to it, and RDF set semantics keep one. That collapse is the entire difference between the .nt file's 42,327,043 lines and the 41,511,580 triples a store loads from it.

This directory is the release the RDF export cannot be: the relational tables themselves, streamed to Parquet (zstd) with every surface form, pipeline status and batch id intact.

Table	Rows	Contents
triples-0000*.parquet (8 shards)	38,450,135	one row per fact: ids and surface labels for subject/predicate/object, object type + statuses, object_description, the 7 pipeline batch ids, created_at
instance_triples.parquet	2,528,556	the instanceOf specialization (entity → concept, + CD/CDG batch ids)
entities.parquet	2,297,995	entity catalog: id, canonical label, description, status, provenance
predicates.parquet	207,633	predicate catalog
concepts.parquet	66,523	concept catalog
batches.parquet	123,444	the provenance containers the batch-id columns point into
manifest.json	—	row counts, shard list, source snapshot

Surface-form variation is preserved as distinct rows: "John Smith, bornIn, NYC" and "Jane Smith, bornIn, Big Apple" are two rows even when both objects resolve to the same E… id. The wording is the disambiguation evidence, and it is the point of the project.

Loading

import duckdb
con = duckdb.connect()
con.sql("SELECT subject_label, predicate_label, object_label, object_id "
        "FROM 'gptkb_v2.0.1_full_record/triples-*.parquet' "
        "WHERE object_id = 'E14' LIMIT 10")

import pandas as pd
entities = pd.read_parquet("gptkb_v2.0.1_full_record/entities.parquet")

All columns are nullable strings, faithful to the source; created_at is the raw pipeline timestamp text.

Earlier versions

The v1 line is archived separately and stays downloadable permanently: GPTKB_v1.5 and GPTKB_v1. Note that v1 used label-based IRIs (/entity/Vannevar_Bush) where v2 uses ID-based ones (/entity/E0); the two are not interchangeable.

Citation

The knowledge base is described in arXiv:2608.03729 and its web interface in arXiv:2608.06992; see gptkb.org/publications for the current references and BibTeX.