ArchT5: A Dataset of Architecture Description Language (ACME and AADL) Snippets Paired with Natural Language Comments

This dataset, ArchT5, has been specifically created to support research on automated comment generation from Architecture Description Language (ADL) snippets.It contains code fragments written in ACME and AADL, paired with their corresponding natural language comments. To facilitate model pretraining and fine-tuning, the dataset also provides derived versions designed for auxiliary tasks: Dual Identifier Tagging (DIT) Dual Masked Identifier Prediction (DMIP) Masked Span Prediction (MSP) The dataset is distributed in JSONL (JSON Lines) format and includes standard training, validation, and test splits.Each entry contains an input field (ADL snippet or masked/tagged snippet) and a target field (corresponding natural language comment or prediction target). This resource is intended to serve as a benchmark for developing and evaluating sequence-to-sequence models such as T5 and CodeT5 in the context of software architecture comprehension.