Hugging Face Overview
AlfredPros/CodeLlama-7b-Instruct-Solidity is a specialized 7 billion parameter Code LLaMA - Instruct model, developed by AlfredPros. It has been fine-tuned specifically for generating Solidity smart contracts from natural language instructions. The model leverages 4-bit QLoRA finetuning, a technique that allows for efficient training and deployment.
Key Capabilities
- Solidity Smart Contract Generation: Translates human-readable instructions into Solidity code for smart contracts.
- Instruction-Following: Designed to understand and execute specific programming tasks based on provided prompts.
- Efficient Training: Utilizes 4-bit QLoRA finetuning, making it accessible for training on more modest hardware (e.g., a single NVIDIA GeForce GTX 1080Ti).
Training Details
The model was trained on the "AlfredPros/smart-contracts-instructions" dataset, which comprises 6,003 pairs of GPT-generated human instructions and corresponding Solidity source code. The training process involved 1 epoch of supervised finetuning, taking approximately 21 hours. Quantization configurations included 4-bit NF4 quantization with double nested quantization, and the optimizer used was paged AdamW 32-bit.