Axolotl

Axolotl

Axolotl is a streamlined, production-grade fine-tuning framework for large language models, released under Apache 2.0. It abstracts away the complexity of LoRA, QLoRA, full fine-tuning, and model parallelism behind a single YAML configuration file, making it the most widely used community fine-tuning toolkit.

Key features

• One YAML config — model, dataset, LoRA rank, quantization, scheduler, all in one place
• Supports LoRA, QLoRA (4-bit), full fine-tune, and ReLoRA
• Multi-GPU and DeepSpeed ZeRO-3 support out of the box
• Compatible with any HuggingFace causal LM (Mistral, Qwen, Phi, Llama, Gemma)
• Flash Attention 2 and xFormers acceleration
• Built-in dataset preprocessing for Alpaca, ShareGPT, Dolly, and custom formats
• Apache-2.0 license

Quick start

pip install axolotl

# Download a sample LoRA config
wget https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/phi/qlora-fft.yml

# Fine-tune (adjust paths and model in YAML first)
axolotl train qlora-fft.yml

# Merge LoRA weights into base model
axolotl merge-lora qlora-fft.yml --lora-model-dir=./outputs/lora

Minimal YAML example

base_model: microsoft/Phi-3-mini-4k-instruct
datasets:
  - path: databricks/databricks-dolly-15k
    type: alpaca
output_dir: ./outputs
adapter: qlora
lora_r: 16
num_epochs: 3

Install via ai-supply

npx ai-supply add axolotl-finetuning

Curated mirror of the open-source Axolotl (Apache-2.0). Get it from the source.