MQuAKE-ST KGQA Dataset quickstart
Static KG · Multi-hop QA · Open dataset

Explore MQuAKE-ST.
Trace the reasoning.

A Wikidata-derived knowledge graph with natural-language questions, explicit starting entities, and two answer settings. Explore real QA records here, or load the dataset into Python.

Explore live questions Load in Python ↓ Browse on Hugging Face ↗
38,516entities
665relations
724,141KG triples
1–4QA hops
Example · Two-hop questionReference path
MacbethQ130283
author · P50
William ShakespeareQ692
citizenship · P27
Kingdom of EnglandQ179876
Path-Key: P50→P272 hops

Dataset reference example, not a MINERVA model prediction.

01 · Live dataset explorer

Browse actual QA records

Inspect questions from the published MQuAKE-ST dataset. Records are fetched on demand through the Hugging Face Dataset Viewer API. No model inference or full KG download is required.

Connecting to Hugging Face Dataset Viewer…

Open source rows ↗

Hugging Face exposes the QA files as API split train. The benchmark partitions are in the row's SplitLabel column.

Question records
Loading questions…
Page 1
Choose a question to inspect its answer and reasoning evidence.

Single-Answer paths shown here are reference evidence, not MINERVA predictions. Multi Answer displays a schematic with unknown intermediate nodes (X₁, X₂, …), not an annotated or reconstructed entity-level path. Property links open Wikidata property pages. This browser queries only QA rows, not all 724,141 graph triples.

02 · Getting started

Load MQuAKE-ST in Python

Download one snapshot of the released dataset, then select the question-answer setting and hop count. Both settings reuse the same KG and metadata.

Terminal · Install once
python -m pip install huggingface_hub pandas
1. Answer setting
2. Hop count
Python · Copy and run
# Loading example appears here.
Single Answer includes one gold answer and annotated reference paths. The mixed-hop file includes 1-hop training questions and 2–4-hop questions.
03 · What's inside

One graph, two QA settings

The dataset uses directed Wikidata identifiers, with separate metadata tables for labels and two alternative question-answer formats.

Full knowledge graph

kg/triplets.txt

Tab-separated head, relation, tail triples with Wikidata QIDs and PIDs. Split graph files are also available.

Single-Answer QA

qa/single_answers/qa_*.csv

One target answer per question, with Paths and Paths-Label annotations.

Multi-Answer QA

qa/multi_answers/qa_*.csv

Lists of valid answers, with Path-Key for relation-chain semantics instead of one gold path.

Wikidata metadata & templates

metadata/*.csv · templates/*.json

Map QIDs and PIDs to readable labels, and inspect verified templates for question relation chains.

QA file names: qa_nhop.csv contains mixed 1–4-hop questions. The hop-specific files qa_2hop.csv, qa_3hop.csv, and qa_4hop.csv can be selected above.
04 · Reading QA records

Read one reference reasoning path

Single Answer stores an ordered path of graph triples. The labeled path is useful for qualitative inspection, while the QID/PID path preserves graph identity.

Illustration from the dataset cardSingle Answer · 2-hop
What is the nationality of the author of Macbeth?
Start: Macbeth (Q130283) · Answer: Kingdom of England (Q179876)
Macbeth— author →William Shakespeare— citizenship →Kingdom of England
Python · Run after loading
# Record inspection example appears here.
05 · Experiment notes

Preserve the evaluation protocol

Keep QA splits separate across answer settings. Single-Answer and Multi-Answer queries may overlap, but the provided train/dev/test assignments are not guaranteed to match. For combined training or cross-setting evaluation, build a shared deduplicated QA split keyed at least by source entity and relation-chain semantics.
1-hop questions are training-only. In the mixed-hop releases, 1-hop examples support learning individual relations. The 2–4-hop questions support multi-hop evaluation.
Directed navigation: Treat the released triples as directed. Adding inverse edges changes the graph and the navigation task. The two answer settings share the exact same static graph.
06 · Continue exploring

Resources

For complete field documentation, pretrained MINERVA models, and experiment details, use the project resources.

MQuAKE-ST dataset card

Full schema, dataset construction, statistics, and attribution.

Open ↗
MINERVA getting-started guide

Dataset preprocessing, training and evaluation configurations.

Guide ↗
THESEUS

Graph navigation benchmark, baselines, model resources.

Project ↗
Theseus in the Graph

The associated research paper.

Paper ↗

MQuAKE-ST-specific materials use CC BY 4.0, subject to applicable upstream terms. Cite MQuAKE, MQuAKE-Remastered, and MQuAKE-ST / THESEUS as listed in the dataset card.