OpenData-Benchmark-ITA is a multiple-choice benchmark dataset designed to evaluate the capability of Large Language Models (LLMs) to understand, retrieve, and reason over public Open Data published by European government portals.

The current release focuses exclusively on Italian Open Data and is based on datasets published on the official Italian government portal, data.gov.it. Future releases will extend the benchmark to include harmonized governmental Open Data from additional European countries, starting with France, Spain, and Germany.

The dataset is released under the ODC-BY (Open Data Commons Attribution) license, enabling broad reuse for research, evaluation, and benchmarking purposes beyond its original project scope.

Benchmark Objective

The primary objective of OpenData-Benchmark-ITA is to assess the effective knowledge and practical usability of Italian governmental Open Data by LLMs developed within the Villanova project.

Rather than testing general language understanding, the benchmark focuses on:

Familiarity with real-world Open Data resources
Ability to interpret dataset metadata
Capability to answer content-based questions grounded in actual public datasets

This makes the benchmark particularly suitable for evaluating domain adaptation, retrieval-augmented generation pipelines, and public-sector–oriented AI systems.

Dataset Composition

The benchmark is structured as a multiple-choice question-answering task. Each question is grounded in the content or metadata of a specific Open Data resource.

Number of datasets sampled: 500
Source portal: data.gov.it
Total datasets available on portal: ~65,000
Data formats: Primarily CSV for data files, paired with JSON metadata

Each benchmark item is derived from a pair consisting of:

A structured data file (mainly CSV)
The corresponding official metadata in JSON format

Data Origin and Curation Process

The dataset is manually curated following a structured and quality-driven workflow.

The process includes:

Systematic sampling from the Italian Open Data portal
Manual verification of dataset relevance and accessibility
Careful inspection and cleaning of metadata
Manual design and validation of multiple-choice questions to ensure clarity, correctness, and grounding in the source data

This approach ensures that the benchmark reflects realistic usage scenarios of public Open Data and avoids synthetic or purely artificial artifacts.

Selection Criteria

The selection of datasets followed clear, content-oriented criteria:

Alignment with the objectives of the Villanova project
Preference for datasets enabling automated processing and analysis
Priority given to machine-readable formats, particularly CSV
Availability of complete and well-structured metadata

The final sample consists of 500 dataset–metadata pairs, each suitable for downstream benchmarking and evaluation tasks.

Collection Period

No strict temporal constraints were applied during dataset selection.

However, preference was given to "live" datasets, identified by:

A recent or regularly updated modification date
Ongoing relevance in terms of data production and maintenance

This choice increases the realism of the benchmark when used to evaluate models intended for interaction with up-to-date public data sources.

Downloads last month: 6