Model components / Fine-tuning code

PEFT

State-of-the-art Parameter-Efficient Fine-Tuning library implementing LoRA, QLoRA, DoRA, adapters, prompt tuning, prefix tuning, and IA3 as a drop-in wrapper around transformers models. Picked because it's the underlying adapter engine that TRL, Axolotl, Unsloth, and most other frameworks compose on top of; touching PEFT is unavoidable in the modern fine-tuning stack. 21.2K GitHub stars, 300 contributors, 116 commits in 90 days.

PEFT (Parameter-Efficient Fine-Tuning) library; repo huggingface/peft, LICENSE confirmed live June 2026. Reference implementation of adapter methods (LoRA, QLoRA, AdaLoRA, IA3, prefix/prompt tuning) integrated across Transformers/Diffusers/Accelerate.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public(huggingface/peft)
governance: Hugging Face
managed-tier: none
core-gated: ungated

Fully OSI-licensed (Apache-2.0), full source public, no proprietary tier.

https://github.com/huggingface/peft/blob/main/LICENSE recorded 2026-06-04
Apache License Version 2.0 text

Adoption

5 high confidence

5.0

Re-scored 4->5. 10.28M PyPI downloads in the last 30 days; the canonical adapter/LoRA library and a dependency of TRL and most LoRA fine-tuning stacks. The map's prevailing usage_volume level-5 floor is >10M/mo (cf. llama-index ~10.18M, pydantic-ai ~31M at level 5); PEFT clears it. Conservative note: at ~10.3M it sits just above the floor, so the bump is borderline and rests on volume plus its dependency-graph centrality in the HF fine-tuning stack.

https://pypistats.org/api/packages/peft/recent recorded 2026-06-25
last_month downloads = 10,281,112

Capability

4 high confidence

4.0

Reference-grade depth on parameter-efficient methods; narrower method scope than full-stack trainers, so C4 not C5.

https://github.com/huggingface/peft/blob/main/LICENSE recorded 2026-06-04
huggingface/peft repo (PEFT adapter library) confirmed live

Unchanged since 2026-07-30 (last edited, not re-checked)