Comparing LLM-based and MDE-based Code Generation for Agile MDE

# OCL to Kotlin & JavaScript Datasets for Code Generation Research

## Overview

This dataset supports research on **automated code generation from OCL (Object Constraint Language)** to two different programming languages: **Kotlin** and **JavaScript**. It was created and used as part of a publication named “Comparing LLM-based and MDE-based Code Generation for Agile MDE” on and is intended for continued research.

It includes:

- Parallel datasets of OCL and target-language code.

- Multiple training subsets for fine-tuning (ranging from ~165 to 1000+ cases), and the displayed datasets are part of the original datasets.

- Predictive metrics of the final fine-tuned models.

- Model behavior demonstration videos after training by using a same test dataset.

---

## Repository Structure

The uploaded files

├── README.md

├── OCLtoKotlinFine-tune/

  ├── 189case/

  ├── 487case/

  └── 1000+case/

└── OCLtoJavascriptFine-tune/

  ├── 165/

  ├── 500+/

  └── 1000+/

---

## Contents

### 1. OCLtoKotlin

- **PartofTrainingDataset/**: Contains aligned OCL and Kotlin code pairs.

- **predictmetrics/**: Files containing accuracy, BLEU score, and other evaluation results.

- **TestModelVideos/**: Recordings demonstrating model performance post-training.

Three different dataset sizes (189, 487, 1000+ cases) were used for iterative training.

---

### 2. OCLtoJavascript

Same structure as above, but targeting JavaScript as the output language.

- Dataset size grows from 165 to 500+ and 1000+ samples.

- Some subsets include **FirstTime** and **SecondTime** training sessions for comparative evaluation by changing the fine-tune parameters using same training dataset.

---

## Usage

This dataset is intended for:

- Training and evaluating code generation models.

- Comparative analysis of model behavior on increasing dataset sizes.

- Display the experiment dataset and results in the associated workshop paper.

---