← Back to Projects

Resource-Efficient & Effective Code Summarization

Saima Afrin, J. Call, K.-N. Nguyen, O. Chaparro, Antonio Mastropaolo

We investigate whether QLoRA (Quantized Low-Rank Adaptation) — a GreenAI parameter-efficient fine-tuning technique proven effective for NL-to-Code tasks such as code generation — can transfer to code summarization, a representative Code-to-NL task. Evaluating two state-of-the-art code language models (CodeLlama and DeepSeek-Coder) on Python and Java, we show that QLoRA enables resource-efficient fine-tuning for summarization while preserving effectiveness, offering a sustainable path for adapting large code models to environments with limited computational resources.