Axolotl

Axolotl turns LLM post-training into a single YAML file: you declare the base model, datasets and prompt formats, method and hardware layout, and that one config carries a run from data prep through evaluation and export. Methods range from full fine-tuning, LoRA and QLoRA to quantization-aware training, DPO, KTO and ORPO preference tuning, GRPO-style reinforcement learning and reward models, across Llama, Qwen, Mistral, Gemma, GPT-OSS and many other Hugging Face architectures, vision and audio included. Scaling features include sample packing, FlashAttention, Liger kernels, sequence and context parallelism, FSDP2, DeepSpeed, expert parallelism for MoE models and multi-node launches, and recent releases added GGUF export for llama.cpp and Ollama. Install with uv on Python 3.12 and a recent PyTorch, or pull the official Docker images; expect an NVIDIA Ampere-or-newer or AMD GPU, and Colab and cloud templates exist. The CLI also prints docs for coding agents. It is Apache-2.0 and free for commercial use, with opt-out telemetry. Aimed at ML engineers who want reproducible fine-tuning without writing training loops.