New Fourier-based technique enables efficient cross-lingual code generation across Python, Java, and other languages using Code Llama 7B

Researchers demonstrate that LoRA fine-tuning on the small MBPP dataset outperforms Code Llama-Python-7B (40.1% vs 38.4% pass@1 score) while being computationally efficient

Study compares Adam and Sophia optimizers, finding Sophia converges faster but with marginal final performance differences

Introduces novel Fourier-based regularization technique to improve parameter-efficient fine-tuning for multilingual code transfer

Addresses enterprise need for cost-effective multi-language support without requiring individual large model fine-tuning for each programming language