ARCHI v2 DATASET

Description Replication package for the article "ARCHI v2: LLM Agents for Service Identification and Architecture Decision Records in Microservice Ecosystems" (M. Calamo, F. Monti, F. Spaziani, D. A. Tamburri, F. Leotta, M. Mecella), submitted to IEEE Transactions on Services Computing. ARCHI v2 is an agentic, LLM-based framework. It turns a system description and its user stories into a microservice architecture, made up of:- services, private data stores, typed inter-service dependencies and the data-management patterns they apply;- a validated PlantUML component diagram;- one Architecture Decision Record (ADR) per design decision, traceable to the user stories. This package contains the generated artefacts, the evaluation results and the analysis scripts behind the article's tables and figures. Contents - ARCHI outputs on 18 student projects from the Archi dataset, for four LLMs (DeepSeek V4.1 Flash, Kimi K2.7-Code, Gemma-4-26B-A4B, Qwen3.8-27B). Each project has architecture.json, a component diagram, a summary and pattern-judge scores. The 160 ADRs come from the DeepSeek run, which also covers PENTION; PENTION's ADRs are not released, so the package contains the ADRs of the 18 student projects.- Zero-shot baseline: diagrams produced by prompting the same models once, plus the prompt used.- MicroserviceDataset: - ARCHI outputs for 19 open-source systems; - the reconstructed requirements documents and reference component diagrams used for scoring.- R2ABench: architecture views generated by ARCHI and by zero-shot prompting with DeepSeek V4.1 Flash.- Modifiability study: - LLM-generated requirement-change scenarios, with the re-derived architectures and per-project reports; - a per-scenario table of the 83 scenarios that all four models re-derived successfully.- Evaluation results: - structural correspondence, rubric scores and smell ratios; - pattern-appropriateness scores from an LLM judge, including a judge that works from the diagrams alone and compares ARCHI with zero-shot prompting on the same evidence; - the R2ABench cross-evaluation.- User study: anonymised questionnaire responses from six practitioners and researchers (P1–P6), with a codebook mapping column codes to the original questions.- Analysis scripts that reproduce the article's figures and aggregates, and the evaluation scripts.