Parameter-Efficient Instruction Tuning Code Large Language Models: An Empirical Study
Abstract: The high cost of full-parameter fine-tuning (FFT) of Large Language Models (LLMs) has led to a series of parameter-efficient fine-tuning (PEFT) methods. However, it remains unclear which methods provide the best cost-performance trade-off at different model scales. We introduce \textsc{Astraios}, a \textbf{fully permissive} suite of 28 instruction-tuned Code LLMs using 7 tuning methods and 4 model sizes up to 16 billion parameters. Through investigations across 5 tasks and 8 different datasets encompassing both code comprehension and code generation tasks, we find that FFT generally leads to the best downstream performance across all scales, and PEFT methods differ significantly in their efficacy based on the model scale. LoRA usually offers the most favorable trade-off between cost and performance. Further investigation into the effects of these methods on both model robustness and code security reveals that larger models tend to demonstrate reduced robustness and less security. Finally, we explore the relationships between updated parameters and task performance. We find that the tuning effectiveness observed in small models generalizes well to larger models, and the validation loss in instruction tuning can be a reliable indicator of overall downstream performance. We believe that our findings of PEFT can generalize to other decoder-only LLMs.\footnote{The codebase (under Apache-2.0 license) and models (under BigCode OpenRAIL-M license) will be publicly available.}
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: Code Generation, Evaluation, Parameter-Efficient Fine-tuning
Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models
Languages Studied: English, Program Languages
Submission Number: 899
Loading