DEFINITION
LoRA adapts large language models by training low-rank matrices beside frozen weights. QLoRA combines 4-bit quantization with LoRA so teams can fine-tune capable models on limited GPU memory.
Why PEFT exists
Full fine-tunes update billions of parameters, demand large GPU fleets, and produce heavy checkpoints. LoRA keeps the base frozen and trains thin adapters that are cheap to store, swap, and version per product line.
QLoRA in practice
QLoRA loads the base in 4-bit precision and trains LoRA adapters in higher precision. It is widely used for domain adapters, instruction styles, and tool-calling habits on single or few GPUs—still requiring careful eval for hallucination and safety.
PEFT vs RAG decision
Fine-tune for stable behavioural gaps; retrieve for facts. If your dataset is mostly documents that change weekly, invest in retrieval and chunking first. If every answer must follow a rigid schema or voice, adapters help.
FAQ
Can I combine LoRA with RAG?
Yes. A common pattern is a LoRA adapter for format/tool use plus RAG for enterprise knowledge and citations.
How much data do I need?
Task-dependent. Many successful adapters use thousands of high-quality examples with clear rubrics; noisy scraped dumps underperform smaller curated sets.
Expert
Deepskilling GenAI Faculty — Programme Lead, Generative AI & LLM tracks. Instructors and mentors for LLM Mastery, fine-tuning, Bedrock/Azure OpenAI delivery, and agentic workflow labs.
Research citations
- LoRA: Low-Rank Adaptation of Large Language Models
Hu, E. et al. (2021). arXiv. Parameter-efficient fine-tuning via low-rank adapters—core to production LLM customization. - QLoRA: Efficient Finetuning of Quantized LLMs
Dettmers, T. et al. (2023). NeurIPS. 4-bit quantization + LoRA for accessible fine-tuning on limited GPUs.