LoRA: Low-Rank Adaptation of Large Language Models
Source: https://arxiv.org/abs/2106.09685
Structured fields
{
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "The foundational paper for LoRA, the adapter method behind most image/video-model style and character fine-tunes. Needed to reason about rank, trainable-parameter counts and why LoRA files are small.",
"source_type": "paper",
"availability": "available",
"source_owner": "Microsoft authors (Hu et al.)",
"next_check_at": "2026-10-24T00:00:00.000Z",
"known_limitations": [
"Evaluates language models; application to diffusion/video models is by later work (see trainer repositories)."
]
}Evidence claims
- [sourced_fact] Paper by Hu et al.; first submitted June 17, 2021, v2 revised October 16, 2021. source
- [upstream_claim] Authors report LoRA freezes pretrained weights and adds trainable low-rank matrices per Transformer layer, matching or outperforming full fine-tuning on RoBERTa, DeBERTa, GPT-2 and GPT-3 while reducing trainable parameters and GPU memory, with no added inference latency (evaluated on language models, not diffusion models). source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | arXiv-nonexclusive-distribution-1.0 | unknown | unknown | unknown | unknown | Page links the arXiv non-exclusive distribution license v1.0 and states no other terms; re-hosting terms for the PDF not analysed. |