Famous Paintings Cultural Heritage Metadata Benchmark for LLM Querying

Famous Paintings Cultural Heritage Metadata Benchmark for LLM Querying

This dataset is a benchmark package built on top of the Famous Paintings dataset from Kaggle and designed for experiments on structured querying over cultural heritage data. It includes the benchmark schema (paintings.json) together with SQL and natural-language query sets (queries_paintings.sql and queries_paintings.txt) for reproducible evaluation. The benchmark models a relational cultural-heritage collection centered on artists, works, museums, subjects, image links, and product-related metadata.

The benchmark includes the following main tables:

• artist
• canvas_size
• image_link
• museum
• museum_hours
• product_size
• subject
• work

Across these tables, the benchmark supports structured information needs involving:

• artist nationality, style, birth year, and death year
• museum country, city, and state
• work style and museum associations
• subject-based retrieval
• image-link retrieval
• sale-price and regular-price filtering
• multi-table joins across artists, works, museums, subjects, and product metadata 

The query set contains 140 manually designed queries in natural-language form, mirrored by corresponding SQL queries, and covers both simple filtering and more complex join-heavy retrieval patterns. These queries were designed to support reproducible experiments on digital libraries, cultural heritage collections, and large language models used as structured knowledge access systems.

License and provenance

This benchmark is derived from the Famous Paintings dataset published on Kaggle (`mexwell/famous-paintings`). At the time of writing, the Kaggle dataset page reports the license as “Other (specified in description)”, not CC0. Reuse of the source data should therefore follow the original dataset terms and any rights information provided by the source publisher.