PEFT means adding small trainable modules while keeping the base LLM frozen. The base weights are never updated together with the adapter.
Four adapter families in LLM-Adapters
- series or bottleneck adapters
- parallel adapters
- prompt-based learning
- reparameterisation methods such as LoRA
Findings from LLM-Adapters
- adapter type and placement are empirical choices across reasoning tasks, and cannot be settled by theory alone
- parallel adapters on the MLP path outperformed the others at a similar budget
- at 7B scale, adapter PEFT with few extra parameters can match or beat full fine-tuning on in-distribution tasks
- LLaMA-13B with a parallel adapter beat ChatGPT on 8 commonsense benchmarks
The reparameterisation family is covered in Low-Rank Adaptation (LoRA).