When prompting falls short but full fine-tuning is too costly, low-rank adaptation (LoRA) offers a smaller weight update. It freezes the pretrained model and trains task-specific adapter matrices, which are combined with the base model at inference. section 5.2 section 5.2.1
For a square weight matrix W of size d × d, LoRA learns two matrices A and B of sizes d × r and r × d. Their product is the additive update ΔW = A × B; the effective weight is W + ΔW. The adapter can be used or removed without changing the base weights. section 5.2.1
Training only the adapter reduces the number of trainable parameters and can make fine-tuning feasible on smaller GPUs. Adapters can also be swapped for different tasks on the same base model. Inference is not necessarily faster: an additional matrix multiplication can increase latency on some hardware. section 5.2 section 5.2.2
Choose and evaluate LoRA as part of an adaptation workflow:
- Establish a prompting baseline on held-out examples before training. If you lack a few hundred high-quality, representative training examples, curate data first. section 1.7
- Prepare message-format data, load the base model, configure and train the adapter, then save it and compare its outputs against the baseline. Hugging Face TRL's SFTTrainer handles adapter wrapping, tokenization, and prompt masking from a LoRA configuration. section 2.8 section 5.1
- Evaluate task quality and test for safety regressions using known harmful prompts. Include explicit refusal examples in training data. section 5.13
Listing 2.3 Step 2, load Qwen3-4B-Instruct-2507 and configure a LoRA adapter
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig
BASE_MODEL = "Qwen/Qwen3-4B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
BASE_MODEL, device_map="auto", dtype="auto", trust_remote_code=True,
)
model.gradient_checkpointing_enable()
model.config.use_cache = False
lora_config = LoraConfig(
r=16, lora_alpha=32, lora_dropout=0.05,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
"up_proj", "gate_proj", "down_proj"],
bias="none", task_type="CAUSAL_LM",
)A configuration for Qwen3-4B shows rank, alpha, dropout, and the attention and feedforward modules targeted by an adapter. section 2.8
IT support house style
An IT support adapter was trained on 166 examples in a fixed house format and compared with minimally prompted and strongly format-prompted base models on six held-out questions. In that small test, only the adapter reached the house format and named the internal tools; the format prompt did not include the internal glossary. section 5.1.8
Use prompting or a decode-time schema when they meet a formatting requirement without training. Use retrieval-augmented generation when knowledge changes frequently or answers need verifiable citations; it can also complement LoRA. If a well-tested adapter plateaus below the quality target, consider full supervised fine-tuning, provided the data and maintenance budget justify it. section 1.6.2 section 5.13 section 6.13
For memory-constrained hardware, QLoRA combines LoRA with a 4-bit base model. LoRA's low-rank constraint also limits capacity: larger datasets or complex tasks may warrant testing full fine-tuning rather than assuming the adapter will close the remaining gap. section 5.13 section 5.2.2
















