🇺🇸 English (EN)

814 Clean Samples: RPGLE -> Java (Clean Arch) | Dataset for AI Training | IBM i Modernization

Curated, verified dataset of 814 RPGLE (Fixed, Free, SQLRPGLE) -> Java translations. Built to train LLMs for IBM i modernization with zero hallucinations. Every sample follows Clean Architecture.

What this solves: Most AI models hallucinate when converting RPGLE because they never saw clean data. This dataset is the antidote.

Stack: Python (processing) + Java 17 / Spring Boot (output) + LoRA + RAG + DB2.

Author: Flavio Frois - IBM i Legacy + Modern AI.

---

🇧🇷 Português (PT-BR) - Original abaixo

🚀 RPG2JAVA - AI-Powered IBM i Modernization Assistant

An AI-powered assistant for IBM i (RPG ILE, CL, DDS) modernization to Java 21 (Clean Architecture).

🛠️ Tech Stack

• LLM: Qwen 2.5 1.5B (Unsloth)
• Fine-Tuning: QLoRA (4-bit)
• RAG: ChromaDB + Sentence Transformers
• Embeddings: all-MiniLM-L6-v2
• Environment: Google Colab (T4 GPU)
• Languages: Python, RPG ILE, CL, Java 21

📊 Dataset

• Total: 814 pares verificados RPGLE -> Java (Clean Arch)
• Antes: 111 Q&A pairs (dataset inicial) -> agora expandido para 814
• Topics: CHAIN, UNLOCK, DATA-GEN, Subfiles, Service Programs, CL calls, IBM i 7.5, Java migration
• Format: JSONL (ChatML)

📂 Project Structure