# Axolotl Axolotl is an open-source fine-tuning framework for [[Large Language Models (LLMs)]]; the practical, batteries-included answer to "I have a model and a dataset, train it." It sits alongside [[Tinker]] as the two primary trainers integrated with [[Atropos]], and is the more popular choice for teams running their own GPUs. The pitch is configuration-as-YAML; describe the model, data, training method, and hardware in a YAML file, run one command. No glue code. ## What it covers - **Full fine-tuning**, **LoRA**, **QLoRA** (quantized LoRA, the dominant practical mode). - **Preference tuning**; DPO, KTO, ORPO. - **Reinforcement learning**; GRPO and friends. - **Reward modeling**. - **Multimodal**; vision-language and audio model training. - **Performance optimizations**; Flash Attention, sequence parallelism, multi-GPU and multi-node training, specialized kernels. - **Data sources**; local, HuggingFace Hub, S3, Azure, GCP, OCI. ## Supported models The open-weight zoo, kept current. Llama (including Llama 4), Mistral, Qwen (including Qwen 3), Gemma, plus MoE architectures and vision-language models. Recent releases ship support for new model families within days. ## How it fits the stack ``` Data + YAML config ↓ Axolotl (or Tinker) ↓ Trained adapter / model ↓ Inference (vLLM, llama.cpp, etc.) or Hermes Agent / Codex CLI / etc. ``` For agent self-improvement specifically, [[Atropos]] generates trajectories, Axolotl or Tinker turns those trajectories into model updates, the updated model goes back into the harness. Axolotl is the "I own the GPUs" path; Tinker is the "I rent them" path. ## License Apache 2.0. Free for commercial and personal use. ## Why it matters Axolotl is the de facto open-source standard for serious fine-tuning. The competition (TRL, Unsloth, Hugging Face Transformers' built-in trainer) covers parts of the surface; Axolotl covers the breadth, with the maturity that comes from broad community use. If you are training or post-training open-weight models on your own infrastructure as of 2026, this is the default choice. ## References - Repository: https://github.com/axolotl-ai-cloud/axolotl ## Related - [[Atropos]] - [[Tinker]] - [[AI Fine-Tuning]] - [[Low Rank Adapter (LoRA)]] - [[AI Open Weight Models]] - [[Reinforcement Learning From Human Feedback (RLHF)]] - [[HuggingFace]]