mAIEnergy Dataset: Multimodal Energy Data for LLM Continual Pre-Training (CPT) and Retrieval‑Augmented Generation (RAG)

The mAIEnergy Dataset is a unified, open-access multimodal corpus designed to support Continual Pre-Training (CPT) of Large Language Models (LLMs) and the development of Retrieval-Augmented Generation (RAG) systems in the energy domain. The dataset integrates heterogeneous energy-related information into structured, machine-readable formats, enabling seamless integration into modern AI and multimodal learning pipelines. The corpus is organised into four top-level modules, namely textual, numerical, geospatial, and imagery, each accompanied by structured metadata and reproducible processing workflows. In total, the dataset contains approximately 50,000 textual documents, 20,000 images, 25 million numerical time-series records, and 2 million geospatial and relational data entries, with a primary focus on the European energy system. 1. Textual (articles‑retrieval) Folder: textual/ Content: Four JSON-Lines (.jsonl) corpora, each storing one document per line: arxiv.jsonl – Scientific preprints and technical papers related to energy systems and AI gov.jsonl – Policy, regulatory, and institutional documents from EU and national government portals news.jsonl – Energy-related news articles retrieved via the GNews API wiki.jsonl – Encyclopedic articles from Wikipedia covering energy technologies, policies, and concepts Preprocessing: HTML and markup removal Title-based deduplication English-language filtering Corrupted or malformed text removal Each document is stored with metadata including title, source URL, document type, and retrieval timestamp. Intended use: Continual pre-training of LLMs on energy-specific policy, technology, and market narratives Construction of dense or hybrid text indices for RAG systems and question-answering applications 2. Numerical (numerical‑retrieval) Folder: numerical/ Structure and content: The numerical module is organised into source-specific subfolders, with CSV files and accompanying JSON metadata: BSO/ – EU Building Stock Observatory building_stock.csv energy_and_environmental_performance.csv financial_performance.csv social_performance.csv reference_buildings.csvEach file is accompanied by a metadata JSON documenting variables, units, geographic coverage, and provenance. ENTSO-E/ – Electricity system time series <country>_actual_generation_<year>.csv <country>_actual_load_<year>.csv <country>_<region>_day_ahead_prices_<year>.csv <country>_installed_capacity_<year>.csvFiles are provided per country, region, and year, with ISO-8601 timestamps and detailed metadata. Eurostat/ – Pan-European statistical indicators annual_energy_balances.csv electricity_prices.csv energy_efficiency_indicators.csv Additional energy, economic, environmental, and socio-economic indicator tables, each with metadata. Open-Meteo/ – Meteorological time series <country>_<city>_<year>.csvDaily aggregated weather variables (temperature, precipitation, wind, radiation, humidity), with metadata describing variables and units. Intended use: Integration of structured numerical signals into hybrid RAG pipelines Training or conditioning models on quantitative energy system features and temporal patterns 3. Geospatial (geospatial‑retrieval) Folder: geospatial/ Subdirectories: cordis/ – CORDIS Horizon 2020 research and innovation networks, represented as node and relationship CSV files (projects, organisations, topics, legal bases) with metadata.json gridkit/ – European high-voltage transmission grid topology (grid_nodes.csv, connected_to_relationships.csv, gridkit.json) osm/ – OpenStreetMap-derived EU energy infrastructure layers (power plants, wind/solar farms, substations, transmission lines, EV charging stations) in GeoJSON, plus neo4j_import/ CSVs for graph ingestion powerplants/ – EU extract of the Global Power Plant Database with nodes, relationships, and eu_powerplants.json tso_network/ – ENTSO-E Transmission System Operator network nodes and interconnection relationships with entsoe_metadata.json Intended use: Spatially grounded RAG combining geospatial vectors with text embeddings Fine-tuning or evaluation of models on spatial reasoning, infrastructure connectivity, and network inference tasks 4. Imagery (images‑retrieval) Folder: imagery/ Providers and content: copernicus/ – Sentinel-2 and Copernicus land-monitoring imagery (JPEG + JSON metadata), organised by region eprel/ – EU energy-label images for appliances (PNG/PDF + JSON metadata) inria/ – High-resolution aerial orthophotos (TIFF + JSON metadata) irf/ – Building façade and industrial imagery (JPEG + metadata) wikimedia/ – Energy-related imagery from Wikimedia Commons with structured metadata Intended use: Training and evaluation of multimodal transformers on energy-related visual content Image-grounded RAG queries combining visual and textual evidence Provenance, metadata, and reproducibility Metadata: Every file is accompanied by structured JSON metadata documenting source attribution, licensing, retrieval timestamps, variable definitions, and processing details. Reproducibility: All data acquisition, cleaning, validation, and structuring steps are implemented as environment-driven, containerised workflows (Docker Compose), enabling full end-to-end reproducibility from raw sources to AI-ready outputs. Scope The mAIEnergy Dataset is intended as a foundational resource for energy-focused AI research, supporting continual pre-training, multimodal retrieval, and decision-support applications. Users may employ the dataset as released or extend it by integrating additional open or proprietary sources using the provided workflows.