Table of Contents
-
Full Fine-Tuning: This approach updates all model parameters during training, enabling maximum model adaptation but requires significant computational resources and storage.
-
Soft Prompting: Instead of altering the model’s core parameters, this method learns a set of “soft prompts” that act as additional input tokens, steering the model towards desired outputs with minimal parameter changes.
-
Adapters: Adapters introduce small, trainable modules into each layer of the model, allowing the main model parameters to remain frozen. This approach is efficient in terms of storage and computation, as only the adapter parameters need updating.
-
Adapter Hub: An extension of the adapter concept, Adapter Hub is a modular framework that enables the sharing and reuse of adapters across tasks. This allows for flexible, plug-and-play fine-tuning of models on multiple tasks with minimal additional training.
-
LoRA (Low-Rank Adaptation): LoRA fine-tunes low-rank matrices within the model, significantly reducing the number of parameters to train. This method is highly parameter-efficient and works well for adapting large language models without substantial resource demands.
Rasoul Zahedifar - [email protected]
GitHub Link: https://github.com/Rasoul-Zahedifar/Parameter-Efficient-Fine-Tunning-of-LLMs