This repository contains an open-source Web 3.0 project leveraging advanced Large Language Models (LLMs) to transform user intents, expressed in natural language, into executable smart contracts. Accelerate your decentralized app development.
Note: The code generated by the LLM is currently of limited quality, often containing “hallucinations” or incomplete logic.
When creating decentralized applications with the smart contract framework, developers regularly need to craft and deploy smart contracts customized to their specific business requirements. This project addresses this challenge by:
-
Accepting a user’s intent in natural language (e.g., “I want an NFT marketplace that charges a commission to sellers”).
-
Processing that intent through a Large Language Model.
-
Generating an initial smart contract draft matching the described functionality.
This project currently supports generating code for the following object types (based on CosmWasm framework and Andromeda Digital Objects):
-
marketplace The Marketplace ADO is a smart contract that allows you to sell your NFTs in a marketplace.
-
cw20 The CW20 ADO is a smart contract to initiate a standard CW20 token. CW20 is a specification for fungible tokens based on CosmWasm.
-
auction The Auction ADO is a smart contract that allows performing custom auctions on NFTs.
-
cw721 The CW721 ADO is a smart contract to allows users to launch their own custom NFT projects.
-
timelock The Timelock ADO or Escrow ADO is a smart contract built to hold funds (Native coins) for a period of time until the set condition is satisfied.
-
cw20 exchange The CW20 Exchange ADO is used to sell CW20 tokens for other assets.
-
splitter The Splitter ADO is a smart contract used to split funds to a preset number of addresses.
-
crowdfund A crowdfund contract that allows users to pledge tokens to reach a funding goal.
To generate test examples:
python create.pyThe scripts are located in the /generated directory
-
Hallucinated / Fantasized Code: The LLM sometimes fabricates or 64FF “hallucinates” contract code that may not compile or make logical sense.
-
Incomplete Edge Cases: Certain flows may be only partially implemented or missing altogether.
-
Deviation from Standards: The generated code may deviate significantly from known, high-quality templates or standard practices.
-
Refined Class Schemas Structuring clearer class models to guide LLM precision.
-
Price/Recall Matrix Establishing robust evaluation metrics based on curated gold-standard datasets.
-
Finetuning the LLM Custom model training with domain-specific data to drastically reduce inaccuracies.
-
Prompt Engineering Systematic, step-by-step prompt structuring to ensure consistent and reliable outputs.
-
Automated Testing Incorporating automated compilation and testing processes, improving deployment reliability.
This project is licensed under the MIT License.