Axolotl is one of the most widely used open-source frameworks for fine-tuning large language models. A single YAML configuration file drives everything from full fine-tuning to LoRA adapters and preference training, across most popular open model families, and the project has about 12,600 stars on GitHub.
Fine-tuning used to mean copying a training script and editing it for every experiment. Axolotl replaced that with a declarative approach: describe the base model, dataset, adapters, optimizer and hardware layout in YAML, and the framework assembles a correct, optimized training run. It supports full fine-tuning as well as LoRA and QLoRA, scales from a single GPU to multi-node clusters with DeepSpeed or FSDP, and incorporates performance techniques such as FlashAttention, sample packing and sequence parallelism. The model coverage is broad โ Llama, Mistral, Qwen, Gemma and many other architectures โ and the training methods extend beyond supervised fine-tuning to preference optimization such as DPO and reinforcement methods such as GRPO, with multimodal training supported as well. Because configurations are files, experiments are reproducible and shareable, which is a large part of why Axolotl became a default choice for open-model post-training.
Axolotl is free and open source under the Apache-2.0 license. Your only costs are the GPUs you train on and any datasets you license.
Axolotl is for ML engineers and researchers who fine-tune open models regularly and want a maintained, configuration-driven framework instead of bespoke training scripts โ from quick LoRA experiments to serious multi-GPU post-training runs.