Axolotl

Free

Open-source framework for fine-tuning and evaluating LLMs. It simplifies the process of experimenting with different training configurations and makes it easy to reproduce and share results, supporting features like LoRA, QLoRA, DeepSpeed, PEFT, and multi-GPU setups.

FreeFree tier

Type

Open Source

Visit Website GitHub (38 stars)

About Axolotl

Axolotl is an open-source framework for fine-tuning large language models (LLMs) with support for a wide range of architectures and training techniques. It simplifies experimenting with configurations like LoRA, QLoRA, DeepSpeed, PEFT, and multi-GPU setups. The framework supports advanced features including FP8 fine-tuning, NVFP4 quantization, MoE LoRA via custom Triton kernels, expert parallelism, and reinforcement learning methods like GRPO. It is actively maintained with frequent support for new models such as Mistral, Qwen, Gemma, and others.

Key Features

LoRA and QLoRA fine-tuning

DeepSpeed and FSDP distributed training

FP8 and NVFP4 quantization support

MoE model fine-tuning with ScatterMoE and SonicMoE

Expert Parallelism for distributed MoE training

Reinforcement learning via GRPO (Async GRPO)

Flash Attention 4 support

Multi-GPU and multi-node training

Wide model support (Mistral, Qwen, Gemma, GLM, etc.)

PEFT integration

Pros & Cons

Pros

Active development with regular updates and new model support
Supports cutting-edge quantization and training techniques
Compatible with many modern LLM architectures
Free and open-source with permissive license (Apache 2.0 inferred)
Extensive documentation and community

Cons

Requires significant GPU memory and compute
Configuration can be complex for newcomers
Dependency on specific hardware (NVIDIA GPUs)

Best For

Fine-tuning open-source LLMs for domain-specific tasksModel alignment through RLHF/GRPOResearch and experimentation with training configurations