What is this?
Axolotl is a free, open-source framework that streamlines post-training and fine-tuning of large language models — you describe your run in one YAML config and it drives the whole pipeline.
Axolotl is a free, open-source framework that streamlines post-training and fine-tuning of large language models — you describe your run in one YAML config and it drives the whole pipeline.Is it a Python library I import, or a tool I run on its own?It's a standalone training tool you run on GPU hardware rather than a library you code against: install it from PyPI or Docker, write a YAML config, and launch training from the command line.
It's a standalone training tool you run on GPU hardware rather than a library you code against: install it from PyPI or Docker, write a YAML config, and launch training from the command line.How do I get my first fine-tune running?Install via Docker (the README calls it the least error-prone route) or pip, start from an example config in the Examples directory, point it at your model and dataset, and follow the Quick Start guide.
Install via Docker (the README calls it the least error-prone route) or pip, start from an example config in the Examples directory, point it at your model and dataset, and follow the Quick Start guide.What kind of machine do I need to run this?Python 3.11 or newer (3.12 recommended), PyTorch 2.11+, and an NVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or an AMD GPU.
Python 3.11 or newer (3.12 recommended), PyTorch 2.11+, and an NVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or an AMD GPU.What's the catch — is any of it hosted for me?No, it's entirely do-it-yourself: there's no hosted service, so you train on your own GPU box or rent one from supported cloud providers like RunPod, Vast.ai or Modal.
No, it's entirely do-it-yourself: there's no hosted service, so you train on your own GPU box or rent one from supported cloud providers like RunPod, Vast.ai or Modal.Does it phone home while I train?Yes, opt-out telemetry is on by default and reports basic system info, model types and error rates — never personal data or file paths. Set AXOLOTLDONOTTRACK=1 to turn it off.
Yes, opt-out telemetry is on by default and reports basic system info, model types and error rates — never personal data or file paths. Set AXOLOTLDONOTTRACK=1 to turn it off.What models and training methods does it support?Models from the Hugging Face Hub — the README names GPT-OSS, LLaMA, Mistral, Mixtral and Pythia, with 2026 additions like Qwen3.5, Gemma 4 and GLM-4.7-Flash. Methods span full fine-tuning, LoRA/QLoRA, QAT (int8/int4/FP8/NVFP4), preference tuning (DPO, IPO, KTO, ORPO), RL (GRPO) and reward modelling, plus vision-language and audio models like Qwen2-VL and Voxtral.
Models from the Hugging Face Hub — the README names GPT-OSS, LLaMA, Mistral, Mixtral and Pythia, with 2026 additions like Qwen3.5, Gemma 4 and GLM-4.7-Flash. Methods span full fine-tuning, LoRA/QLoRA, QAT (int8/int4/FP8/NVFP4), preference tuning (DPO, IPO, KTO, ORPO), RL (GRPO) and reward modelling, plus vision-language and audio models like Qwen2-VL and Voxtral.Do I have to write much code to configure a run?No — one YAML configuration file drives the entire pipeline: dataset preprocessing, training, evaluation, quantization and inference. You mostly pick an example config and edit settings rather than write training code.
No — one YAML configuration file drives the entire pipeline: dataset preprocessing, training, evaluation, quantization and inference. You mostly pick an example config and edit settings rather than write training code.Where can my training datasets live?Almost anywhere: Axolotl loads datasets from your local disk, the Hugging Face Hub, or cloud storage on S3, Azure, GCP and OCI.
Almost anywhere: Axolotl loads datasets from your local disk, the Hugging Face Hub, or cloud storage on S3, Azure, GCP and OCI.What if a single GPU isn't enough for my model?It scales across hardware: multi-GPU training via FSDP1/FSDP2 or DeepSpeed, multi-node runs with Torchrun or Ray, plus parallelism tricks like sequence and context parallelism for long-context training.
It scales across hardware: multi-GPU training via FSDP1/FSDP2 or DeepSpeed, multi-node runs with Torchrun or Ray, plus parallelism tricks like sequence and context parallelism for long-context training.Which topics is it in?distributed-training, moe, lora, training, reinforcement-learning, quantization, fine-tuning, multimodal, huggingface, pytorch, deep-learning, framework
distributed-training, moe, lora, training, reinforcement-learning, quantization, fine-tuning, multimodal, huggingface, pytorch, deep-learning, framework