Axolotl Fine-Tuning Framework Docs
by Axolotl AI
Post-train an open-weight LLM from a single YAML file - SFT, DPO, GRPO and multi-node included.
Overview
Axolotl is the configuration-driven fine-tuning framework maintained by Axolotl AI, and docs.axolotl.ai is the reference an engineer actually works from rather than a course to sit through. The entire workflow is one YAML file plus five CLI verbs: 'axolotl fetch examples' to pull starter configs, then preprocess, train, inference and merge-lora. The docs are organised as Getting Started (Quickstart, Installation, Inference/Merging), Model Guides for 40+ families including Llama 4, Mistral, Qwen 3, Gemma, Phi, Granite, Jamba and multimodal or audio models such as Voxtral, Dataset Formats (streaming pre-training versus 'completion', template-free 'input_output', 'chat_template' conversation covering ChatML, OpenAI messages, ShareGPT and custom Jinja2 templates, plus Alpaca-style instruction data), Deployments (Docker, multi-GPU, multi-node, Ray Train, AMD HPC, Mac M-series), How-To Guides (multimodal, RLHF, GRPO, reward modelling, quantization), Core Concepts (batch sizing, preprocessing, mixed precision, attention types), Advanced Features (FSDP, PyTorch ao, sequence parallelism, N-D parallelism) and Troubleshooting including a dedicated NCCL debugging page. Supported methods span full fine-tuning, LoRA, QLoRA, GPTQ, QAT, preference tuning (DPO, IPO, KTO, ORPO), reinforcement learning (GRPO, GDPO) and reward or process-reward modelling. Requirements are Python 3.11+ and PyTorch 2.11+ on an Ampere-or-newer NVIDIA GPU, an AMD GPU, or Mac M-series silicon. The project is Apache-2.0, started April 2023, carries roughly 12.3k GitHub stars with commits landing daily, and lists 170+ contributors and users including Nous Research, Modal, Replicate and OpenPipe.
At a Glance
- Topic
- Fine-Tuning
- Level
- Advanced
- Format
- Documentation
- Cost
- Free
- Duration
- ~2-3 hours to install and finish a first LoRA run, then ongoing reference
- Provider
- Axolotl AI
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Drive an entire fine-tuning run from a single declarative YAML config file
- ✓Choose between full fine-tuning, LoRA, QLoRA and QAT against your GPU budget
- ✓Format datasets as Alpaca, ChatML, OpenAI messages, ShareGPT or template-free input_output
- ✓Run preference tuning with DPO, IPO, KTO and ORPO on your own pairs
- ✓Set up GRPO reinforcement learning and reward or process-reward modelling pipelines
- ✓Scale past one GPU using FSDP, DeepSpeed ZeRO, Ray Train and sequence parallelism
- ✓Merge LoRA adapters back into base weights and run inference on the result
- ✓Tune micro-batch size, sample packing and mixed precision to fit VRAM limits
Highlights
- •Config-as-code: the YAML is the experiment record, so runs are reproducible and reviewable in git
- •One config system spans SFT, DPO/IPO/KTO/ORPO, GRPO and reward modelling - no separate tool per method
- •40+ documented model guides, so a new architecture usually has a known-good starting config
- •Multi-GPU and multi-node are first-class: FSDP, DeepSpeed ZeRO, sequence and N-D parallelism
- •Apache-2.0 with 170+ contributors and production users including Nous Research, Modal and Replicate
- •Candid tradeoff: an independent 2026 framework comparison found single-GPU QLoRA wall-clock trails Unsloth, and warns that YAML misconfigurations can fail quietly
Who It’s For
Best For
- ✓ML engineers fine-tuning open-weight models on their own GPUs
- ✓Teams running multi-GPU or multi-node training who need working FSDP/DeepSpeed configs
- ✓Practitioners building RLHF or GRPO alignment pipelines without writing a trainer
- ✓Anyone who wants reproducible training runs checked into version control
Prerequisites
- •Solid Python and PyTorch - this is a framework reference, not an introductory course
- •Python 3.11 or newer (3.12 recommended) with PyTorch 2.11 or later installed
- •An Ampere-or-newer NVIDIA GPU, an AMD GPU, or a Mac M-series machine
- •Understanding of what LoRA, quantization and a training epoch actually are
FAQ
What is Axolotl Fine-Tuning Framework Docs?
The official documentation for Axolotl, the YAML-configured fine-tuning framework used to post-train open-weight LLMs. It covers installation, dataset formats, LoRA/QLoRA/full fine-tuning, preference tuning and GRPO reinforcement learning, plus multi-GPU and multi-node deployment. Written for engineers who already know PyTorch, it gets you from a fresh GPU box to a merged, working fine-tuned model driven entirely by a single config file you can check into git.
Is Axolotl Fine-Tuning Framework Docs free?
Axolotl Fine-Tuning Framework Docs is free to access.
What level is Axolotl Fine-Tuning Framework Docs for?
Axolotl Fine-Tuning Framework Docs is aimed at a advanced audience. Recommended background: Solid Python and PyTorch - this is a framework reference, not an introductory course, Python 3.11 or newer (3.12 recommended) with PyTorch 2.11 or later installed, An Ampere-or-newer NVIDIA GPU, an AMD GPU, or a Mac M-series machine, Understanding of what LoRA, quantization and a training epoch actually are.
How long does Axolotl Fine-Tuning Framework Docs take?
Expect roughly ~2-3 hours to install and finish a first LoRA run, then ongoing reference. Most learners work through it at their own pace.
What will I learn from Axolotl Fine-Tuning Framework Docs?
You'll learn: Drive an entire fine-tuning run from a single declarative YAML config file; Choose between full fine-tuning, LoRA, QLoRA and QAT against your GPU budget; Format datasets as Alpaca, ChatML, OpenAI messages, ShareGPT or template-free input_output; Run preference tuning with DPO, IPO, KTO and ORPO on your own pairs; Set up GRPO reinforcement learning and reward or process-reward modelling pipelines; Scale past one GPU using FSDP, DeepSpeed ZeRO, Ray Train and sequence parallelism; Merge LoRA adapters back into base weights and run inference on the result; Tune micro-batch size, sample packing and mixed precision to fit VRAM limits.
Topics
Sources
This page was written from 6 sources, 3 on domains other than docs.axolotl.ai.