How to fine-tune Llama 3.1 (LoRA, QLoRA)

See the Llama 3 hub for sizes. Parameter-efficient methods (LoRA and QLoRA) are the usual way to adapt these models on a single GPU. Estimate cost with the training cost calculator or read the fine-tuning overview.

Which checkpoint

• Start from the Instruct checkpoint to adjust style, format or domain while keeping chat behavior. Meta describes it as tuned with supervised fine-tuning and RLHF.
• Start from the base checkpoint for large-scale or heavily different instruction data, where you supply the whole post-training recipe.

The cards document no LoRA target modules, so choose them with your training library's defaults for Llama architectures.

Chat template

Train with the same template inference uses: format examples with the tokenizer's apply_chat_template. The Instruct card says tool use is supported through that template, so tool-calling data should be rendered the same way.

License limits on derivatives

From the Llama 3.1 Community License text:

• If you distribute a model trained or improved using Llama materials, include "Llama" at the beginning of its name.
• Prominently display "Built with Llama" on a related website, interface, blog post, about page or product documentation.
• Keep the notice "Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved" in a Notice file.
• Licensees with more than 700 million monthly active users must request a license from Meta.
• Use must follow Meta's Acceptable Use Policy.

Other notes

• Official languages are eight (English, German, French, Italian, Portuguese, Hindi, Spanish, Thai). Fine-tuning data in other languages needs its own evaluation.