Feb–Mar 2024
Llama-2 Fine-Tuning
An early exercise in learning parameter-efficient fine-tuning by doing it: LoRA on the Platypus dataset. The notebook is public.
Stack
Python, Llama-2, LoRA, PEFT, Jupyter
Links
Notes
This was mostly about learning LoRA by using it rather than reading about it: fine-tune Llama-2-7b-chat on Platypus, then check whether target-task accuracy actually moved.
The numbers
- Method
- LoRA and PEFT supervised fine-tuning on the Platypus dataset.
- Result
- Target-task accuracy improved from 73% to 87%.
- Notebook
- Public supervised fine-tuning and evaluation workflow.