LPFGen: Smart Contract Generation with Prompt-Guided and Fine-Tuned LLMs

Smart contracts enable autonomous blockchain applications but remain difficult to implement securely and efficiently. We present LPFGen, a prompt–plus–fine-tuning framework for automated contract synthesis with large language models (LLMs). Our approach leverages Qwen3-8B with LoRA adapters, trained on a mixed corpus of 141,000 contract–prompt pairs and 6,003 masked-function repair samples, providing a reproducible and efficient PEFT recipe. Evaluation across 100 user-style specifications combines usability screening, compilation success, and static analysis, along with held-out repair and detection tasks. Results show that LPFGen raises the proportion of usable generations from 58.3% to 78.2%, while also improving detection and repair performance (F₁ from 0.376 to 0.613; Accuracy from 0.696 to 0.787). These consistent gains indicate clear improvements in generation quality and reliability over untuned baselines. By integrating prompt design with targeted fine-tuning, LPFGen demonstrates that principled adaptation of LLMs can yield secure, verifiable, and practically deployable smart contracts. All code and experiment scripts are released to support reproducibility and further research.