Model components / Fine-tuning code

Axolotl

YAML-config-driven fine-tuning framework, wraps TRL, PEFT, DeepSpeed, FSDP, and bitsandbytes behind a single config file so users can run SFT, LoRA, QLoRA, full fine-tunes, and DPO across 100+ model families without writing training loops. Picked for serious multi-GPU production fine-tuning where reproducibility via config matters and where Unsloth's single-GPU optimizations don't apply. 12K GitHub stars, 240 contributors, 185 commits in 90 days; the de facto tool for hobbyist and indie fine-tuners shipping models on HF Hub.

Axolotl fine-tuning framework (axolotl-ai-cloud/axolotl), confirmed live June 2026 (Apache-2.0). Config-driven post-training: full FT, LoRA/QLoRA/GPTQ/QAT, DPO/IPO/KTO/ORPO, GRPO/GDPO, reward modeling; multi-GPU FSDP1/FSDP2/DeepSpeed, multi-node via Torchrun/Ray, plus SP and ND parallelism.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public(axolotl-ai-cloud/axolotl)
no managed SaaS tier (cloud-template integrations only: RunPod/Modal/etc.)

Fully OSI-licensed (Apache-2.0); no proprietary managed training service, only deployment templates for third-party clouds.

https://github.com/axolotl-ai-cloud/axolotl/blob/main/LICENSE recorded 2026-06-04
Apache License Version 2.0 text

Adoption

2 medium confidence

2.0

~22.3K PyPI downloads last month; ~12K GitHub stars. Popular config-driven FT tool in the open-model community but download volume is modest vs TRL/PEFT/Unsloth; level 2 on usage_volume, stars corroborate only.

https://pypistats.org/packages/axolotl recorded 2026-06-04
Downloads last month: 22,336
https://github.com/axolotl-ai-cloud/axolotl recorded 2026-06-04
~12,000+ GitHub stars (corroborating, not basis)

Capability

4 high confidence

4.0

Very wide method and parallelism coverage (ND parallelism, multi-node), strong C4; below Megatron-LM frontier-scale-training anchor on demonstrated extreme scale.

https://github.com/axolotl-ai-cloud/axolotl recorded 2026-06-04
documented methods (LoRA/QLoRA/DPO/GRPO/...) and parallelism (FSDP/DeepSpeed/SP/ND, multi-node Ray)

Unchanged since 2026-06-24 (last edited, not re-checked)