Axolotl
A tool for fine-tuning LLMs with a focus on ease of use and flexibility. Axolotl supports various fine-tuning methods (full, LoRA, QLoRA), multiple model architectures, and different dataset formats. It provides YAML-based configuration for reproducible training runs.
Implements
Concepts this tool claims to implement:
- Fine-Tuning primary
Full fine-tuning and parameter-efficient methods (LoRA, QLoRA, ReLoRA). Support for Llama, Mistral, Mixtral, Falcon, and other architectures. Flash attention, gradient checkpointing, DeepSpeed integration.
- Instruction Tuning primary
Dataset formats for instruction tuning (Alpaca, ShareGPT, completion). Chat template support. Multi-turn conversation training.
- RLHF secondary
DPO (Direct Preference Optimization) support. Reward model training capabilities. RLHF-style alignment methods.
Integration Surfaces
Details
- Vendor
- OpenAccess AI Collective (community)
- License
- Apache-2.0
- Runs On
- local, cloud
- Used By
- human, system
Links
Notes
Axolotl is the go-to tool for fine-tuning open LLMs. Balances power and usability - YAML configs make runs reproducible while supporting advanced features. Active community with frequent updates for new models and methods. Often used with runpod or Lambda Labs for GPU access.