ASTRAIOS is a suite of 28 instruction-tuned OctoCoder models using 7 tuning methods and 4 model sizes up to 16 billion parameters. The study investigates the effectiveness of parameter-efficient fine-tuning (PEFT) methods for code large language models (Code LLMs) across 5 tasks and 8 datasets. The results show that full-parameter fine-tuning (FFT) generally provides the best downstream performance, while PEFT methods vary in effectiveness based on model scale. LoRA typically offers the best cost-performance trade-off. Larger models tend to have reduced robustness and security. The study also explores the relationship between updated parameters, cross-entropy loss, and task performance, finding that the effectiveness of small models generalizes to larger ones, and validation loss can predict downstream performance. The ASTRAIOS suite includes models fine-tuned with various PEFT methods, including adapter-based, prompt-based, and intrinsic-rank-based tuning. The study highlights the importance of model robustness and code security, showing that larger models are more vulnerable to adversarial examples and generate less secure code. The results suggest that while FFT remains the best for overall performance, PEFT methods like LoRA and Parallel Adapter are competitive, especially for smaller models. The study also finds that the performance of PEFT methods can be predicted by cross-entropy loss, which is a strong indicator of downstream performance. The findings contribute to understanding the effectiveness of different PEFT methods for Code LLMs and their implications for real-world applications.ASTRAIOS is a suite of 28 instruction-tuned OctoCoder models using 7 tuning methods and 4 model sizes up to 16 billion parameters. The study investigates the effectiveness of parameter-efficient fine-tuning (PEFT) methods for code large language models (Code LLMs) across 5 tasks and 8 datasets. The results show that full-parameter fine-tuning (FFT) generally provides the best downstream performance, while PEFT methods vary in effectiveness based on model scale. LoRA typically offers the best cost-performance trade-off. Larger models tend to have reduced robustness and security. The study also explores the relationship between updated parameters, cross-entropy loss, and task performance, finding that the effectiveness of small models generalizes to larger ones, and validation loss can predict downstream performance. The ASTRAIOS suite includes models fine-tuned with various PEFT methods, including adapter-based, prompt-based, and intrinsic-rank-based tuning. The study highlights the importance of model robustness and code security, showing that larger models are more vulnerable to adversarial examples and generate less secure code. The results suggest that while FFT remains the best for overall performance, PEFT methods like LoRA and Parallel Adapter are competitive, especially for smaller models. The study also finds that the performance of PEFT methods can be predicted by cross-entropy loss, which is a strong indicator of downstream performance. The findings contribute to understanding the effectiveness of different PEFT methods for Code LLMs and their implications for real-world applications.