Base Llama knows everything on the internet but nothing about your company, your data, or your specific task. Fine-tuning teaches it your domain.
Full fine-tuning changes all 8 billion parameters. That needs multiple A100 GPUs and costs thousands. LoRA (Low-Rank Adaptation) changes less than 1% of parameters and works on a single GPU.
Rex — Fine-Tune Llama with LoRA
Base Llama knows everything on the internet but nothing about your company, your data, or your specific task. Fine-tuning teaches it your domain.
Full fine-tuning changes all 8 billion parameters. That needs multiple A100 GPUs and costs thousands. LoRA (Low-Rank Adaptation) changes less than 1% of parameters and works on a single GPU.