The challenge will use openly licensed source materials covering transport, buildings and energy, and urban greenery. These will be organised into case packs containing the information needed to answer a set of practical questions.

Materials may include reports and technical guidance, spreadsheets, asset registers, maps, spatial datasets, scans and plans. Each pack includes source identifiers, dates, provenance and licence information. Read each Question together with its Context, which sets the relevant snapshot, assumptions and calculation conventions.

Participants will receive:

1. Development examples: Train.csv contains questions with checked reference answers and supporting evidence.
2. Test questions and source materials: Test.csv contains the complete set of Test questions and source references, without reference answers. Submit an answer and supporting evidence for every Test ID, including questions reserved for final evaluation. SourceMaterials.zip contains the documents, tables, maps and other materials used by Train and Test. Extract it beside the CSV files into a sources folder, or let the starter notebook do this. SourceManifest.csv records source IDs, filenames, publishers, licences, attribution and file hashes.
3. A submission template: SampleSubmission.md defines the required Markdown format and does not contain reference answers. Include every Test ID exactly once, followed by an answer in natural English and evidence identifying the original source pages, cells, records, features or model elements. Invalid file structure is rejected; a structurally valid blank answer scores zero.
4. A starter notebook: StarterNotebook.ipynb provides examples of loading the data, exploring source materials, generating an answer with a small model and preparing a submission.

Questions are based on real, openly licensed source materials and may include clearly labelled hypothetical scenarios. Reference answers are independently checked and kept private, together with the public/private assignment of test cases. Each source retains its original licence and attribution requirements.

Train and Test materials are separated by related source groups. Test questions, Test sources, their public copies and derivatives, and generated answers must not be used for training, fine-tuning, distillation, self-training or pseudo-labelling. Processing Test sources at inference time for OCR, parsing, indexing, retrieval, spatial analysis and calculations is permitted without training or adapting models on them.

Participants may use the provided datasets and approved external datasets listed in AllowedDatasets.csv, subject to their original licences and the challenge rules. The register distinguishes training/development sources from Test sources, which are inference-only.

Additional external datasets may be permitted after organiser review. To propose or declare an additional dataset, please email the organisers with the subject line “GenAI for Good Challenge: Dataset proposal” by 30 November 2026. Include the dataset link, version, licence and intended use, and obtain approval before using it. Approved additions will be listed in AllowedDatasets.csv and announced on this Data page, with updates on Mondays when new datasets are approved. All sources, versions and licences used in your solution must also be declared during code review.

More information.