Research map / Architectures and efficiency / Quantization
QLoRA: Efficient Finetuning of Quantized LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| NF4+DQ instead of FP4 quantization (ablation) | QLoRA · MMLU 5-shot (LLaMA 7-65B) | accuracy: 52.2 -> 53.1 | Helped |
| 4-bit NF4+DQ QLoRA instead of 16-bit LoRA | LLaMA 7-65B finetuning · MMLU 5-shot | accuracy: 53.0 -> 53.1 | No clear effect |
| NF4 data type instead of Int4 (ablation) | 4-bit quantized LLMs · Pile Common Crawl | perplexity: 34.34 -> 27.41 | Helped |
| double quantization (ablation) | QLoRA · zero-shot / LM | accuracy | No clear effect |
| QLoRA NF4+DQ vs full BF16 finetuning | T5-3B · Super-NaturalInstructions | RougeL: 54.3 -> 55.3 | No clear effect |
| apply LoRA to all linear layers (ablation) | LoRA on LLaMA-7B · Alpaca / MMLU | accuracy | Helped |
| vary LoRA rank r (ablation) | QLoRA · Alpaca | performance | No clear effect |
| train on target only instead of instruction+response (ablation) | QLoRA LLaMA-7B · MMLU 5-shot | accuracy: 37.5 -> 38.6 | Helped |
| increase dataset size 50k -> 150k (ablation) | QLoRA · MMLU 5-shot | accuracy: 39.28 -> 40.02 | No clear effect |
| finetune on OASST1 instead of FLAN v2 | LLaMA-65B QLoRA · Vicuna benchmark | % of ChatGPT: 48.4 -> 99.3 | Helped |
| finetune on FLAN v2 instead of OASST1 | LLaMA-7B QLoRA · MMLU 5-shot | accuracy: 36.6 -> 44.5 | Helped |
Closest papers by meaning
- VeLoRA: Memory Efficient Training using Rank-1 Sub-Token Projections (NeurIPS 2024 · similarity 0.90)
- LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models (ICLR 2024 · similarity 0.89)
- BitDelta: Your Fine-Tune May Only Be Worth One Bit (NeurIPS 2024 · similarity 0.88)
- LoQT: Low-Rank Adapters for Quantized Pretraining (NeurIPS 2024 · similarity 0.88)
- FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic (NeurIPS 2025 · similarity 0.88)
- LowRA: Accurate and Efficient LoRA Fine-Tuning of LLMs under 2 Bits (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.