Axolotl is a free, open-source LLM fine-tuning framework designed to simplify the post-training and fine-tuning process of the latest large language models (LLMs). It is designed around YAML configuration files, allowing users to complete dataset preprocessing, training or fine-tuning, model inference, and evaluation within the same configuration.
Main capabilities
- Multi-model support: Can train a variety of models on the Hugging Face Hub, including GPT-OSS, LLaMA, Mistral, Mixtral, Pythia, and more.
- Multimodal training: Supports fine-tuning vision-language models (VLMs), including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, GLM-4.6V, InternVL 3.5, Gemma 3n, PaddleOCR-VL, Muse Glimmer, as well as audio models such as Voxtral, with support for images, video, and audio.
- Training methods: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT (int8/int4/FP8/NVFP4/MXFP4), FP8 mixed-precision training, NVFP4/MXFP4 MoE LoRA, preference tuning (DPO, IPO, KTO, ORPO), reinforcement learning (GRPO, GDPO), reward modeling (RM)/process reward modeling (PRM).
- Configuration approach: A single YAML configuration file can cover the complete fine-tuning workflow of dataset preprocessing, training, evaluation, quantization, and inference.
- Performance optimization: Multipacking, Flash Attention 2/3/4, Xformers, Flex Attention, SageAttention, Liger Kernel, Cut Cross Entropy, ScatterMoE, sequence parallelism (SP), LoRA optimization, multi-GPU training (FSDP2, DeepSpeed), multi-node training (Torchrun, Ray), etc.
- Dataset handling: Can load datasets from local, HuggingFace, and cloud storage (S3, Azure, GCP, OCI).
- Cloud-ready: Provides Docker images and PyPI packages, usable on cloud platforms and local hardware.
Usage Quick start requires an NVIDIA GPU (Ampere or newer, for bf16 and Flash Attention) or AMD GPU, Python >=3.12 (3.12 recommended), PyTorch ≥2.13.0 (2.13.0 recommended). Installing with Docker is usually less error-prone than installing in your own environment; see the official documentation for other installation methods. The official Getting Started Guide is provided as a more detailed introductory guide.
Documentation and support The official documentation covers installation options, support matrix, configuration guide, dataset loading, dataset guide, multi-GPU training, multi-node training, API reference, and FAQ. Axolotl includes built-in documentation optimized for AI coding agents (Claude Code, Cursor, Copilot, etc.), distributed with the pip package, with no need to clone the repository; if using the source repository, the agent documentation is located at docs/agents/, and the project overview is in AGENTS.md. You can also join the Discord community, view the Examples directory, and check the Debugging Guide. For dedicated support, contact [email protected].
Other information Axolotl enables opt-out telemetry by default to understand project usage and prioritize improvements, collecting basic system information, model types, and error rates, without collecting personal data or file paths; it can be disabled by setting AXOLOTL_DO_NOT_TRACK=1. The project uses the Apache 2.0 license.
