Fine-Tune LLMs with LLaMA Factory

Basic idea

LLaMA Factory is an open-source framework that simplifies training and fine-tuning large language models. It offers a unified interface for methods such as supervised fine-tuning (SFT), RLHF, and QLoRA, and supports a wide range of LLM architectures including LLaMA, Mistral, and Qwen.

• Unified CLI — one interface for supervised, RLHF, and parameter-efficient training
• Broad model support — works with popular open LLM families and Hugging Face workflows
• Flexible fine-tuning — LoRA, QLoRA, and full fine-tuning with ready example configs

What you'll accomplish

You'll set up LLaMA Factory on your hardware platform to fine-tune large language models with LoRA, QLoRA, and full fine-tuning methods, using the CLI and example configs from the upstream repository.

What to know before starting

Required:

• Basic Python knowledge for editing config files and troubleshooting
• Command-line usage for shell commands and virtual environments
• Familiarity with PyTorch and the Hugging Face Transformers ecosystem
• Fine-tuning concepts: tradeoffs between LoRA, QLoRA, and full fine-tuning

Optional:

• Dataset preparation: formatting text data into JSON for instruction tuning
• Resource management: adjusting batch size and memory settings for GPU constraints

Supported hardware platforms

Use the matrix below to confirm your hardware platform, recommended default local settings, and whether multi-node applies.

Prerequisites

Hardware requirements

• Supported hardware platform — see Supported hardware platforms matrix above
• Sufficient storage space (plan for >50 GB for models and checkpoints): df -h

Software requirements

• CUDA 12.9 or newer: nvcc --version
• Git: git --version
• Python 3 with venv and pip: python3 --version && pip3 --version
• Internet connection for downloading models from Hugging Face Hub

Ancillary files

All required assets are in the LLaMA Factory repository.

• examples/train_lora/qwen3_lora_sft.yaml — example LoRA SFT training configuration
• examples/inference/qwen3_lora_sft.yaml — example chat/inference configuration for the fine-tuned adapter
• examples/merge_lora/qwen3_lora_sft.yaml — example export/merge configuration for production use
• Data preparation docs — dataset formatting guidance

Time & risk

• Estimated time: 60 MIN for initial setup; training can take 1–7 hours depending on model size and dataset
• Risk level: Medium
  • Model downloads require significant bandwidth and storage
  • Training may consume substantial GPU memory and need batch-size or accumulation tuning
• Rollback: Deactivate the virtual environment and remove the factoryEnv and LLaMA-Factory directories. Delete local training checkpoints to reclaim storage.
• Last Updated: 07/31/2026
  • Set up LLaMA Factory with a venv-based PyTorch CUDA 13 workflow for Qwen3 LoRA fine-tuning, chat validation, and export