🧩 AI FrameworksOpen SourceFree

Axolotl

YAML-configurable fine-tuning for LoRA, QLoRA, and full fine-tuning

4.4/ 5⭐ 12,000 GitHub starsPythonApache-2.0

📋 Overview

Axolotl is an open-source fine-tuning framework that wraps multiple approaches (LoRA, QLoRA, full fine-tuning, DPO, RLHF) in a simple YAML-configurable interface. Unlike Unsloth (which focuses on speed), Axolotl prioritizes flexibility — supporting 50+ model architectures (Llama, Mistral, Gemma, Qwen, Phi, Falcon, MPT), multiple datasets formats, and advanced techniques like flash attention, gradient checkpointing, and distributed training. The YAML config makes experiments reproducible and shareable, and the active community maintains templates for most popular models.

✨ Key Features

•12K+ GitHub stars — most flexible fine-tuning framework
•YAML-configurable: model, dataset, method in one file
•50+ model architectures supported
•LoRA, QLoRA, full fine-tuning, DPO, RLHF
•Flash attention 2, gradient checkpointing, FSDP
•Built-in GGUF export for local inference
•Active community with model-specific templates

🎯 The Problem It Solves

Fine-tuning LLMs requires stitching together multiple libraries (PEFT, bitsandbytes, transformers, datasets) with careful memory management. Axolotl collapses this complexity into a single YAML config — define your model, dataset, and method, then run.

🔧 How It Works

You write a YAML config specifying base model, dataset, fine-tuning method (LoRA/QLoRA/full), and training parameters. Axolotl handles the rest: model loading with quantization, dataset preprocessing, training with HuggingFace Trainer, and adapter merging. It supports flash attention 2, gradient checkpointing, and distributed training via FSDP. The framework auto-detects your hardware and adjusts settings. Export to GGUF for local inference is built-in.

🚀 Installation & Quick Start

Installation

See website

Quick Start

See documentation

✅ Pros

•Most flexible fine-tuning framework
•YAML config makes experiments reproducible
•Broadest model support (50+ architectures)
•Multiple methods: LoRA, QLoRA, DPO, RLHF
•Built-in GGUF export
•Apache 2.0 licensed — completely free
•Active community with templates

❌ Cons

•Not as fast or memory-efficient as Unsloth
•Requires more GPU RAM for equivalent models
•YAML config has a learning curve
•Documentation assumes ML expertise
•Breaking changes between versions

💬 Practitioner Verdict

“Axolotl is the most flexible fine-tuning framework — the YAML config makes experiments reproducible, and the model support is unmatched. The trade-off: not as fast or memory-efficient as Unsloth, requires more GPU RAM for equivalent models, and the YAML config has a learning curve. For researchers and teams who need maximum flexibility and model choice, Axolotl is the default.”