axolotl-ai-cloud/axolotl

★ 11,895PythonAudience · researcherComplexity · 5/5Setup · hard

Mindmap

mindmap
  root((axolotl))
    What it does
      Fine-tune LLMs
      YAML config driven
      No custom code needed
    Supported Models
      Llama
      Mistral
      Qwen
      Gemma
    Fine-tuning Methods
      LoRA
      QLoRA
      Full fine-tuning
    Distributed Training
      Multi-GPU
      DeepSpeed
      FSDP

mindmap root((axolotl)) What it does Fine-tune LLMs YAML config driven No custom code needed Supported Models Llama Mistral Qwen Gemma Fine-tuning Methods LoRA QLoRA Full fine-tuning Distributed Training Multi-GPU DeepSpeed FSDP

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Fine-tune a Llama or Mistral model on a custom instruction dataset to make it follow domain-specific commands using LoRA to keep GPU memory low.

USE CASE 2

Train a specialized chatbot for a vertical domain by fine-tuning an open-source base model on curated question-answer pairs with a YAML config.

USE CASE 3

Run distributed fine-tuning across multiple GPUs using DeepSpeed or FSDP to handle models too large for a single card.

USE CASE 4

Experiment quickly with different fine-tuning approaches by swapping base model or technique in one YAML file without rewriting training code.

Tech stack

PythonPyTorchCUDAYAMLDeepSpeedFSDP

Getting it running

Difficulty · hard Time to first run · 1day+

Requires one or more NVIDIA GPUs with CUDA. CPU-only or MPS (Apple Silicon) training is not the intended path.

In plain English

axolotl-ai-cloud/axolotl is a framework for fine-tuning large language models, which are the AI systems behind chat tools and text generators. Fine-tuning means taking an existing pre-trained model and training it further on a specific dataset so it learns to behave differently or focus on a particular domain. Axolotl makes this process more accessible by handling much of the technical setup through configuration files. You define what you want to do in a YAML file, a plain-text settings file, by specifying which base model to start from, what training data to use, and which fine-tuning technique to apply. Axolotl then runs the training process. This approach lets practitioners experiment without writing new training code from scratch each time. The project supports a wide range of popular AI models including Llama, Mistral, Qwen, Gemma, and many others. The README contains a detailed changelog listing dozens of model additions over the past year, and new models are added regularly with each update. It also supports multiple fine-tuning methods, including LoRA (a technique that reduces how much memory and compute is needed compared to full fine-tuning). For people training on multiple GPUs or across multiple machines, Axolotl includes support for distributed training approaches like FSDP and DeepSpeed, which split the workload across hardware to handle models that would not fit on a single GPU. It also supports quantization, a technique that reduces model size and memory usage by storing weights in lower precision number formats. The README is detailed and technically oriented, aimed at ML practitioners who already have some familiarity with the model training workflow. The project is open-source and free to use.

Copy-paste prompts

Prompt 1

Write an Axolotl YAML config file to fine-tune Llama-3-8B on a custom instruction dataset stored as a JSONL file using LoRA with rank 16, training for 3 epochs on a single A100 GPU.

Prompt 2

I want to fine-tune Mistral-7B using Axolotl with QLoRA (4-bit quantization + LoRA) to fit in 24GB VRAM. Show me the quantization and LoRA config fields I need in the YAML.

Prompt 3

How do I set up Axolotl to run distributed fine-tuning across 4 GPUs using DeepSpeed ZeRO-3? Show me the relevant config block and the launch command.

Prompt 4

My Axolotl training run is producing loss spikes after 500 steps. What YAML config settings for learning rate schedule, gradient clipping, or warmup should I check first?

Open on GitHub → Explain another repo

← axolotl-ai-cloud on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.