The Open-Source Models Beating GPT-4 on Coding Benchmarks