Axolotl

A free fine-tuning framework driven by one YAML file, with LoRA, QLoRA, preference tuning and NVFP4 and expert-parallel training for MoE models.

  • Fine-tuning
  • Apache 2.0 · Open source
  • Linux
  • Verified

Axolotl is a free and open-source tool for post-training and fine-tuning large language models. One YAML configuration file covers dataset preprocessing, training, evaluation, quantization and inference. The training methods listed in the README include full fine-tuning, LoRA, QLoRA, quantization-aware training (int8, int4, FP8, NVFP4 and MXFP4), preference tuning (DPO, IPO, KTO and ORPO), reinforcement learning (GRPO and GDPO), and multimodal training for vision-language and audio models.

The README's update log tracks new open models soon after release, including Gemma 4, Qwen3.6, Mistral Large 4, Reflection's Beam, MiniCPM5 and Muse Glimmer, with expert-parallel and NVFP4 LoRA training for mixture-of-experts models. It supports distributed training with FSDP2, DeepSpeed and context parallelism, and it can export a model to GGUF with axolotl export for use with llama.cpp and Ollama.

The README lists an NVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or an AMD GPU, Python 3.12 or newer and a recent PyTorch as requirements, and it provides Docker images and PyPI packages. The license is Apache 2.0, and a fine-tuned model still carries the license of the base model you started from.

Good for

  • Repeatable fine-tuning runs from a config file
  • LoRA and QLoRA on one GPU, or multi-GPU and multi-node training
  • Fine-tuning new mixture-of-experts models with expert-parallel and NVFP4 LoRA training
  • Exporting a fine-tuned model to GGUF

Watch out for

  • It needs an NVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or an AMD GPU
  • The README asks for Python 3.12 or newer and a recent PyTorch, so use a fresh environment or the Docker image
  • A fine-tuned model inherits the license of its base model, including any revenue limits

Links