Axolotl
YAML-config LLM fine-tuning for serious trainers: QLoRA to distributed MoE.
About
Axolotl is the free, open-source fine-tuning framework for people who train models seriously: declare everything in a YAML config and get QLoRA/LoRA, full fine-tuning, FSDP2, DeepSpeed, expert parallelism for MoE models, context parallelism, and one-command GGUF export for llama.cpp/Ollama deployment. It tracks new architectures aggressively (Qwen3.6/3.8, MiniCPM5, NVFP4 LoRA in late 2026) and ships Docker images as the primary install. More powerful and more complex than Unsloth — choose Axolotl when you need distributed training and fine-grained control. Apache-2.0, 12.6k stars.
What it can do
- **YAML-config training**: reproducible configs for every technique
- **Efficient methods**: LoRA, QLoRA, DoRA, RSLoRA, plus full fine-tuning
- **Distributed**: FSDP2, DeepSpeed, expert/context/tensor parallelism for MoE
- **Quantized training**: NVFP4 LoRA, QLoRA with CPU-staged loading
- **Export**: GGUF, merged fp16, and adapter formats for deployment
- **Cutting-edge coverage**: new architectures supported within weeks of release
Setup
```yaml
# config.yaml
base_model: Qwen/Qwen2.5-7B-Instruct
adapter: qlora
sequence_len: 4096
datasets:
- path: data/train.jsonl
type: chat_template
```
```bash
# via Docker (recommended)
axolotl train config.yaml
axolotl export config.yaml # GGUF for llama.cpp/Ollama
```
Docs: docs.axolotl.ai. Colab notebooks available for single-GPU runs.
Prerequisites
NVIDIA GPU(s) — single consumer GPU for QLoRA, multi-GPU for full/distributed. Docker (recommended) or Python 3.10+ with CUDA.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 12,557 GitHub stars (log-scaled vs catalog peers); blended with ~8,818 monthly package downloads; last push 0d ago; 11 releases/12mo; 247 open issues. Manual review: power-user fine-tuning (YAML configs, QLoRA/FSDP2/MoE, GGUF export); extremely active; steeper learning curve than Unsloth; PyPI understates usage (Docker primary). License: Apache-2.0.
★ 12,557 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.2/10
- Maintenance · recency, releases, issue load
- 8.6/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · 12,557 GitHub stars (log-scaled vs catalog peers)
- · blended with ~8,818 monthly package downloads
- · last push 0d ago; 11 releases/12mo; 247 open issues
- · manual review: power-user fine-tuning (YAML configs, QLoRA/FSDP2/MoE, GGUF export); extremely active; steeper learning curve than Unsloth; PyPI understates usage (Docker primary)
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.