Smart contract generation model based on code annotation and AST-LSTM tuning

Abstract

With the wide application of smart contracts in many fields, the number, types, and complexity of smart contracts are showing a rapidly increasing trend. However , the development of smart contracts has its own unique programming language and security requirements, which are difficult for conventional software personnel to adapt to quickly, and how to realize the efficient development of smart contracts according to the application requirements is an important issue that needs to be solved for its further development. To this end, this paper proposes a smart contract generation method based on AST-LSTM characterization and code annotation tuning large language model, which adopts AST-LSTM model combining Abstract Syntax Tree (AST) and Tree-LSTM to vectorize the code as well as Sentence-Bert to vectorize the annotations and carry out a weighted analysis, and constructs a smart contract clustering analysis model to achieve accurate clustering of functionally similar smart contracts. Then the AST-LSTM+Transformer model is used to detect defects in the clustered code and correlate the related annotation information to construct a diverse Prompt feature prompt statement dataset. Finally, the LLaMA2-7B model is used as the basis for demand-specific smart contract generation with the help of Lora and P-Tuning v2 fine-tuning techniques. In this paper, with the help of BLEU, an auxiliary tool for bilingual translation quality assessment, and Mythril, VaaS, 1 and other code security detection tools, we conducted comparative experiments with existing methods. The results of the experiment show that the average value of BLEU of the code generated by this paper’s method is improved by about 25%, and the code security is improved by about 9%, which will greatly promote the rapid development and exploitation of smart contracts with high-security requirements.