ArchPipeline — ADL Comprehension Dataset (ACME/AADL)

This dataset supports the comprehension phase of ArchPipeline, a two-phase transfer-learning framework for software architecture comprehension and evolution (under review, 2026). Each instance pairs a formal ADL (Architecture Description Language) code fragment with its natural language description, enabling sequence-to-sequence models to learn the mapping from formal architectural syntax to human-readable documentation. The dataset covers two ADL formalisms: ACME (an academic architectural interchange language) and AADL (Architecture Analysis and Design Language, used in safety-critical embedded systems). STATISTICS- Full corpus : 1,092 pairs (400 ACME + 692 AADL before balancing)- Train : 640 pairs (320 ACME + 320 AADL)- Validation : 80 pairs (40 ACME + 40 AADL)- Test (raw) : 80 pairs (40 ACME + 40 AADL)- Test (Jaccard-filtered) : 66 pairs (33 ACME + 33 AADL) The Jaccard-filtered test set retains only instances with mean Jaccard distance ≥ 0.35 from the training set and quality score ≥ 4, ensuring evaluation reflects genuine generalization rather than surface-level pattern memorization. DATA FORMATEach line is a JSON object with the following fields:- input : formal ADL code fragment (ACME or AADL)- target : natural language description- adl_type : ADL formalism (ACME or AADL)- element_type: architectural element category (15 categories)- quality : quality score 0–6- source : source document identifier CONSTRUCTIONAll fragments were manually collected from peer-reviewed scientific publications on ACME and AADL. The dataset was processed through a seven-stage pipeline covering noise removal, text normalization, ADL type annotation, element annotation, quality scoring, quality filtering, and stratified 80/10/10 splitting with seed 42. AADL augmentation was applied using four strategies to address formalism imbalance. FILES- train.jsonl : 640 training pairs- val.jsonl : 80 validation pairs- test.jsonl : 80 raw test pairs- test_v2.jsonl : 66 Jaccard-filtered test pairs (used in all experiments)- dataset_full.jsonl: full corpus of 1,092 pairs RELATED RESOURCES- Evolution dataset : https://doi.org/10.5281/zenodo.20497909- Code and notebooks: github.com/lamia-aissaoui/ArchPipeline