Axolotl

Axolotl is a free, open-source LLM fine-tuning framework for developers and researchers adapting models to their own data. It supports training on local hardware or cloud infrastructure you control.

• Fine-tuning methods: Use full fine-tuning, LoRA or QLoRA, alongside quantization-aware training and mixed-precision methods.
• Text and multimodal models: Train families including GPT-OSS, Llama, Mistral, Qwen and Gemma, plus vision-language and audio models such as Qwen2-VL, Pixtral, LLaVA and Voxtral.
• Preference and reward training: Apply preference methods such as DPO, reinforcement learning with GRPO, and reward or process reward modelling.
• Larger training workloads: Distribute training across multiple GPUs or machines using FSDP, DeepSpeed, Torchrun or Ray, with sequence packing and attention optimizations to improve hardware use.
• Datasets and evaluation: Use local, Hugging Face or cloud datasets, with a shared configuration covering preprocessing, training, evaluation, quantization and inference.

Axolotl uses the Apache 2.0 license and runs in Docker and Kubernetes environments as well as on local hardware. It requires an NVIDIA or AMD GPU; some precision and attention features require NVIDIA Ampere or newer hardware. Local datasets don't have to be uploaded to an external AI service.

Telemetry is enabled by default and can be disabled. It collects basic system information, model types and error rates, excluding personal data and file paths.