Which statement is true about Fine-tuning and Parameter-Efficient Fine-Tuning (PEFT)?

Fine-tuning involves updating all the weights of a pre-trained model using a new, task-specific dataset. This process is computationally intensive and requires significant memory as it modifies the entire model. In contrast, Parameter-Efficient Fine-Tuning (PEFT) freezes the vast majority of the pre-trained model's parameters and introduces a small number of new, trainable parameters or modifies a small subset of existing ones. This approach drastically reduces the computational and storage costs associated with training, making it possible to adapt large models with much less data and fewer hardware resources. PEFT achieves performance comparable to full fine-tuning on many tasks while being significantly more efficient.

B. PEFT modifies or adds a small set of parameters to the existing architecture; it does not replace the entire architecture and is significantly less data-intensive.
C. Both techniques adapt a pre-trained model. Neither involves training from scratch, which would be far more computationally expensive than even full fine-tuning.
D. Both methods involve model parameter modification. The core difference is the number of parameters modified, not the type of data used for training.

1. Oracle Official Documentation: Oracle Cloud Infrastructure (OCI) Documentation, "Generative AI, Fine-tuning Models". The documentation states, "Fine-tuning custom models in the Generative AI service uses a technique called Parameter-Efficient Fine-Tuning (PEFT). With PEFT, instead of retraining all the parameters in the model, only a small number of parameters are updated." This directly supports the efficiency and minimal parameter update claims in option A.

2. Peer-Reviewed Academic Publication: Hu, E. J., Shen, Y., Wallis, P., et al. (2021). "LoRA: Low-rank adaptation of large language models." International Conference on Learning Representations. In the abstract, the authors state, "We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices... LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times." This quantifies the core principle of PEFT described in option A. (Available at: https://arxiv.org/abs/2106.09685)

3. Peer-Reviewed Academic Publication: Lialin, V., Deshpande, V., & Rumshisky, A. (2023). "Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning." Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Section 1 (Introduction) contrasts full fine-tuning, which "updates all of the model’s parameters," with PEFT, which "freezes the pre-trained LLM and adds a small number of trainable parameters." This reinforces the distinction made in option A. (DOI: https://doi.org/10.18653/v1/2023.acl-long.465)