Guide to fine-tuning Llama models on your own data in the cloud

The guide recommends starting with the smallest Llama Instruct model that can handle the task, and suggests Llama 3.1 8B Instruct as a default for most support, extraction and house-style work. Before training, users must accept Meta's license on Hugging Face under the same account as their token, since a token alone does not grant access and approval is per repository. Examples should be formatted with the exact Llama 3 chat template used at serving time, because the Base and Instruct variants differ in their chat formats and end tokens. Memory needs are estimated from parameter counts, with LoRA suited to most sizes and QLoRA needed only for the 70B model on a single 80 GB GPU. The authors also note that fine-tuning does not meaningfully update the December 2023 knowledge cutoff, so fresh facts should be supplied through retrieval instead.