We provide a simple finetuning commands (litgpt finetune_*) that instruction-finetune a pretrained model on datasets such as Alpaca, Dolly, and others. For more information on the supported instruction datasets and how to prepare your own custom datasets, please see the tutorials/prepare_dataset tutorials.
LitGPT currently supports the following finetuning methods:
litgpt finetune_full
litgpt finetune_lora
litgpt finetune_adapter
litgpt finetune_adapter_v2The following section provides more details about these methods, including links for additional resources.
The section below provides additional information on the available and links to further resources.
litgpt finetune_fullThis method trains all model weight parameters and is the most memory-intensive finetuning technique in LitGPT.
More information and resources:
- the LitGPT tutorials/finetune_full tutorial
litgpt finetune_lora stabilityai/stablelm-base-alpha-3bLoRA and QLoRA are parameter-efficient finetuning technique that only require updating a small number of parameters, which makes this a more memory-efficienty alternative to full finetuning.
More information and resources:
- the LitGPT tutorials/finetune_lora tutorial
- the LoRA paper by (Hu et al. 2021)
- the conceptual tutorial Parameter-Efficient LLM Finetuning With Low-Rank Adaptation (LoRA)
litgpt finetune_adapter stabilityai/stablelm-base-alpha-3bor
litgpt finetune_adapter_v2 stabilityai/stablelm-base-alpha-3bSimilar to LoRA, adapter finetuning is a parameter-efficient finetuning technique that only requires training a small subset of weight parameters, making this finetuning method more memory-efficient than full-parameter finetuning.
More information and resources:
- the LitGPT tutorials/finetune_adapter tutorial
- the Llama-Adapter (Gao et al. 2023) and Llama-Adapter v2 (Zhang et al. 2023) papers that originally introduces these methods
- the conceptual tutorial Understanding Parameter-Efficient Finetuning of Large Language Models: From Prefix Tuning to LLaMA-Adapters