Scalable Code Generation with Large Language Models: An Open-Source Ensemble and Reinforcement Learning–Based Selector

Replication Package: (Mohammed A. Shehab*, Safwan Omari, Mohammad Wardat, and Yaser Jararweh – Accepted) This repository provides the replication materials for the study “Scalable Code Generation with Large Language Models: An Open-Source Ensemble and Reinforcement Learning–Based Selector”. It includes datasets, experiment scripts, and evaluation pipelines used to compare commercial large language models (Codex, Copilot, PaLM2) with two novel open-source methods: Open-Source Ensemble — synthesizes multiple LLM outputs to improve correctness. RL-Based Model Selector — applies reinforcement learning to dynamically select efficient model subsets. 📊 Contents dataset/ – programming tasks, evaluation logs, model outputs, and metrics. ensemble_LLMs.py – ensemble synthesis pipeline. train_rl_weights.py / RL_route_torch.py – reinforcement learning agent training and routing. offline_eval_torch.py / eval_resynthesis_torch.py – offline evaluation utilities. Dataset Information.xlsx – benchmark metadata and summary. ⚙️ Environment Tested with Python 3.9+, PyTorch, Transformers, Gymnasium, Lizard, NumPy, Pandas, and Matplotlib.The requirements.txt file is included for reproducibility. 🔬 Purpose The package supports transparency and independent replication of our experiments on accuracy, code quality, runtime performance, and efficiency (token cost + latency). ⚖️ License Released under the MIT License. All datasets are anonymized and contain no personal or proprietary information.