anonymous-7989/herikgqa-benchmark

HeriKGQA: benchmark and dataset snapshot

A 500-question benchmark for question answering over a cultural-heritage sensor knowledge graph, and the anonymised dataset snapshot the knowledge graph is built on. The code (system, Ontop mapping, ontology, scorer) is in the accompanying code repository, linked from the paper.

Files

File	Contents
benchmark_500.json	the benchmark, with typed reference values and conversation history (use this file for evaluation)
benchmark_500.csv	the same questions and reference answers as text (for browsing)
benchmark_500.meta.json	counts, data range and the conventions the references follow
snapshot.dump	the dataset: PostgreSQL 16, custom format (pg_restore)

Benchmark

500 questions, 100 of each type:

Type	Asks about
basic	counts, averages and extremes per site and quantity, comparisons between sites
spatial	distances between nodes and between sites, nearest nodes, nodes within a radius, centroids, altitudes, northern- and southernmost nodes
temporal	first and latest readings, time spans, daily and monthly statistics, values in date ranges and relative windows, when an extreme occurred
anomaly	whether a quantity at a site exceeds its expert (AD-RT) rule thresholds on a day or at the latest reading, and the thresholds themselves
follow-up	a question that depends on one or two earlier turns (another site, the next day or month, the minimum instead of the maximum, …)

Every type includes a few unanswerable questions: the answer must say that the data cannot answer them.

Each item of benchmark_500.json has:

Field	Meaning
id, type, template	identifier, question type, question template
question	the question
history	earlier turns of a follow-up (question, answer)
reference	the reference answer, as text
reference_value	the value an answer must contain (see below)
answerable	false when the data cannot answer the question; the answer must say so

reference_value is typed:

Type	An answer is correct when
integer	a number in it equals the count
float	a number in it equals the value at the stated precision, within 1%
Yes / No / Cannot be determined	it gives that anomaly verdict
date or YYYY-MM-DD HH:MM UTC	it states that date or time
site or node name	it names that site or node
list	every element matches
null	it declines (no data), not a number

Every reference is computed directly from the snapshot with SQL and plain Python, independently of any QA system.

The conventions:

• all times are UTC, and "on a date" means that UTC day;
• relative times ("the last 24 hours") end at the snapshot's last reading;
• distances are haversine on a 6,371,000 m sphere;
• an anomaly verdict is Yes if any applicable rule fires, No if at least one rule was evaluated and none fired, and Cannot be determined if none could be evaluated.

They are given in full in benchmark_500.meta.json.

Dataset snapshot

A frozen copy of the sensor dataset, 13 May – 12 August 2026: 3,737,747 observations at five heritage sites, plus the expert anomaly rules (thresholds, windows, quantities).

It is anonymised:

• sites are named site_a … site_e;
• place names in node, rule and threat descriptions are replaced;
• each site is moved to a fictitious location, keeping the distances between its nodes.

Measured values, timestamps and row counts are unchanged.

To restore it:

createdb herikg_snapshot
pg_restore -d herikg_snapshot --no-owner --no-privileges snapshot.dump

The code repository's Dockerfile downloads this dataset and serves the snapshot as a virtual knowledge graph through Ontop.