Which statement accurately reflects the differences between these approaches in terms of the number of parameters modified and the type of data used?

Full fine-tuning involves updating all the weights of a pre-trained model using a labeled, task-specific dataset. This process is computationally intensive and creates a full copy of the model for each task. In contrast, Parameter-Efficient Fine-Tuning (PEFT) methods, such as LoRA (Low-Rank Adaptation), freeze the original model's weights and only train a small number of new or existing parameters. This approach significantly reduces computational and storage costs while still effectively adapting the model to a new task using the same kind of labeled, task-specific data.

A. Continuous pretraining uses unlabeled, domain-specific data to adapt the model's general knowledge, not labeled, task-specific data.

B. Parameter-Efficient Fine-Tuning and Soft Prompting modify only a small subset of parameters (or add new ones) and use labeled data, not all parameters with unlabeled data.

D. Continuous pretraining modifies all of the model's original parameters, contrary to the claim that it requires no modification.

1. Oracle Cloud Infrastructure (OCI) Documentation, Generative AI, "Custom models and fine-tuning": "Fine-tuning updates all of the weights of a pretrained base model... This method is also known as full fine-tuning." and "Parameter-Efficient Fine-Tuning (PEFT) is a method that updates only a small subset of the model's parameters... PEFT is faster and less computationally expensive than full fine-tuning." This source directly supports the distinction made in option C.

2. Lialin, V., Deshpande, V., & Rumshisky, A. (2023). Scaling Down to Scale Up: A Guide to Parameter-Efficient Fine-Tuning. Section 2, "Background". arXiv:2303.15647. This paper defines full fine-tuning as updating all parameters of a pre-trained model for a downstream task. It then introduces PEFT methods as an alternative that tunes "only a small number of (extra) model parameters," which aligns perfectly with the description in option C.

3. Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., ... & Chen, W. (2021). LoRA: Low-Rank Adaptation of Large Language Models. Section 3, "Our Method". arXiv:2106.09685. The foundational paper for LoRA, a popular PEFT technique, states: "We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture." This confirms that PEFT methods add and train a few new parameters while keeping the original model frozen.