axolotl-ai-cloud/axolotl

What it solves

Axolotl simplifies the complex process of post-training and fine-tuning large language models (LLMs). It provides a unified framework that removes the need to write custom training scripts for every new model or technique, allowing users to configure their entire pipeline—from dataset preprocessing to quantization—via a single YAML file.

How it works

Axolotl acts as a wrapper and orchestrator for various training libraries and optimizations. It supports a wide array of models (such as Llama, Mistral, and Qwen) and modalities (text, vision, audio). Users define their training parameters in a configuration file, and Axolotl handles the underlying technical execution, including memory optimizations like Flash Attention and distributed training across multiple GPUs or nodes using FSDP and DeepSpeed.

Who it’s for

It is designed for AI researchers, developers, and organizations who want to fine-tune open-source LLMs on custom datasets without managing the boilerplate code associated with different training frameworks.

Highlights

• Broad Model Support: Compatible with a vast range of models from the Hugging Face Hub, including multimodal vision-language and audio models.
• ** converges multiple techniques**: Supports full fine-tuning, LoRA, QLoRA, Preference Tuning (DPO, ORPO), RL (GRPO), and Reward Modelling.
• Advanced Optimizations: Integrates Flash Attention 2/3/4, Sequence Parallelism, and various quantization methods (FP8, NVFP4) to reduce VRAM usage and increase speed.
• Cofiguration-Driven: Uses a single YAML file to manage the entire pipeline, making experiments reproducible and easy to share.
• Agent-Ready: Includes built-in documentation specifically optimized for AI coding agents like Cursor and Claude Code.