Fine-Tuning a Small LLM to Generate Solidity Tests

The Testing Gap

Smart contracts handle real money. Yet most projects have inadequate test coverage — not because developers are lazy, but because writing comprehensive fuzz tests is hard.

The Approach

I fine-tuned Mistral 7B on 1,847 Solidity contracts paired with their Foundry test suites.

Building the Dataset

Scraped from OpenZeppelin, Uniswap V3, Aave V3, Compound V2, and 1,200+ audited contracts from Code4rena.

Filtered aggressively:

• Only contracts with > 80% coverage
• Only test files with fuzz tests
• Only passing suites

Fine-Tuning with LoRA

from peft import LoraConfig, get_peft_model
 
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
    task_type="CAUSAL_LM"
)
 
model = get_peft_model(model, lora_config)
# Trainable params: 13.6M (0.19% of 7.24B)

Training: 4 hours on a single A100. Cost: ~$16.

Results

Example Output

function testFuzz_TransferExceedsBalance(uint256 amount) public {
    vm.assume(amount > 0 && amount <= type(uint128).max);
    token.mint(alice, amount - 1);
 
    vm.prank(alice);
    vm.expectRevert("ERC20: transfer amount exceeds balance");
    token.transfer(bob, amount);
}

Real, useful tests — not boilerplate.

What's Next

Building a VS Code extension that suggests tests as you write Solidity — running locally via Ollama, no API costs.