跳到正文
arXiv cs.AI· Terry Yue Zhuo, Armel Zebaze, Nitchakarn Suppattarachai, Leandro von Werra, Harm de Vries, Qian Liu, Niklas Muennighoff·· 4 小时前AI 评分33

Astraios:代码大语言模型的参数高效指令微调研究

Astraios: Parameter-Efficient Instruction Tuning Code Large Language Models

AI 导读

Astraios 是一套包含 28 个指令微调 OctoCoder 模型的套件,涵盖 7 种微调方法和最高 16B 参数的 4 种模型规模,在 5 个任务和 8 个数据集上评估代码理解与生成。

正文 · 原文

View PDF HTML (experimental)

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 Astraios, a suite of 28 instruction-tuned OctoCoder models 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. At last, we explore the relationships among updated parameters, cross-entropy loss, 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.
Comments: 25 pages (12 main), 19 figures, 8 tables
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
Cite as: arXiv:2401.00788 [cs.CL]
  (or arXiv:2401.00788v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2401.00788

arXiv-issued DOI via DataCite

Submission history

From: Terry Yue Zhuo [view email]
[v1] Mon, 1 Jan 2024 15:30:19 UTC (2,123 KB)
[v2] Fri, 13 Nov 2026 11:35:07 UTC (2,113 KB)

来源:arXiv cs.AI · arxiv.org