security-auditor-grpo/README.md

---
base_model: Qwen/Qwen2.5-Coder-0.5B-Instruct
library_name: transformers
model_name: security-auditor-grpo
tags:
- generated_from_trainer
- grpo
- trl
- security
- smart-contracts
- solidity
- audit
- web3
license: apache-2.0
datasets:
- oxdev/smart-contract-security-sft
- oxdev/smart-contract-security-audit-v2
pipeline_tag: text-generation
language:
- en
---

# 🔐 Smart Contract Security Auditor (GRPO)

A specialized **smart contract security auditor** built on [Qwen2.5-Coder-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct), fine-tuned using **Group Relative Policy Optimization (GRPO)** on real-world audit findings from top security firms.

## 🎯 What It Does

Given a Solidity smart contract, this model identifies security vulnerabilities and produces structured audit findings with:
- Vulnerability classification (reentrancy, access control, oracle manipulation, etc.)
- Severity assessment (Critical/High/Medium/Low)
- Detailed description of the vulnerability
- Impact analysis
- Proof of concept exploit code
- Recommended fixes

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model = AutoModelForCausalLM.from_pretrained(
    "oxdev/security-auditor-grpo",
    use_cache=True,  # Important: config has use_cache=False from training
)
tokenizer = AutoTokenizer.from_pretrained("oxdev/security-auditor-grpo")
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device="cuda")

messages = [
    {"role": "system", "content": "You are an expert smart contract security auditor. Analyze the provided Solidity code for vulnerabilities."},
    {"role": "user", "content": """Audit this contract:
```solidity
contract SimpleBank {
    mapping(address => uint256) public balances;
    function deposit() public payable { balances[msg.sender] += msg.value; }
    function withdraw(uint256 amount) public {
        require(balances[msg.sender] >= amount);
        (bool success, ) = msg.sender.call{value: amount}("");
        require(success);
        balances[msg.sender] -= amount;
    }
}
```"""},
]

result = pipe(messages, max_new_tokens=512, do_sample=False, return_full_text=False)
output = result[0]["generated_text"]
if isinstance(output, list):
    output = output[-1]["content"]
print(output)
```

## 🔗 Try It Live

**Interactive Demo:** [oxdev/security-auditor-demo](https://huggingface.co/spaces/oxdev/security-auditor-demo) — Side-by-side comparison with base model, 7 test cases with known vulnerabilities, automated scoring.

## Training Details

### V1 (Current Model)
- **Method:** GRPO (Group Relative Policy Optimization)
- **Base Model:** Qwen2.5-Coder-0.5B-Instruct
- **Dataset:** [oxdev/smart-contract-security-sft](https://huggingface.co/datasets/oxdev/smart-contract-security-sft) (327 synthetic samples)
- **Hardware:** NVIDIA T4 (16GB)
- **Epochs:** 2
- **Reward Functions:** Format compliance, finding rate
- **Results:**
  - Format reward: 0.025 → 0.40 (**16× improvement**)
  - Finding rate: 0% → 50-75%
  - Mean reward: -0.34 → -0.006

### V2 (Pending — Colab Notebook Ready)
- **Dataset:** [oxdev/smart-contract-security-audit-v2](https://huggingface.co/datasets/oxdev/smart-contract-security-audit-v2) (50,902 real audit findings)
- **Sources:** SkywardNomad92/smart-contract-audit-findings, samscrack/cyfrin-audit-findings, Solodit API
- **4 Reward Functions:** Format (0.25), Severity matching (0.25), Category matching (0.25), Quality (0.25)
- **Train on Colab:** Open [`train_grpo_v2_colab.ipynb`](https://huggingface.co/oxdev/security-auditor-grpo/blob/main/train_grpo_v2_colab.ipynb) in Google Colab with a free T4 GPU

## Vulnerability Categories Covered

| Category | Keywords |
|----------|----------|
| Reentrancy | reentrancy, reentrant, callback |
| Access Control | unauthorized, permission, onlyowner |
| Oracle Manipulation | price feed, chainlink, twap |
| Flash Loan | flash loan, flashloan |
| Overflow/Underflow | overflow, underflow, arithmetic |
| Front-running | front-run, sandwich, MEV |
| DoS | denial of service, gas limit, unbounded |
| Token Issues | fee-on-transfer, rebasing, ERC20 |
| Storage | storage collision, delegatecall, proxy |
| Cross-chain | bridge, relay, message passing |
| Liquidation | liquidation, collateral, health factor |
| Signature | ecrecover, replay, nonce, EIP712 |
| Initialization | uninitialized, constructor |
| Rounding | precision, truncation, decimal |

## Architecture

- **Model:** Qwen2ForCausalLM
- **Parameters:** 0.5B
- **Hidden Size:** 896
- **Layers:** 24
- **Attention Heads:** 14 (2 KV heads)
- **Context Length:** 32,768 tokens
- **Chat Template:** ChatML (`<|im_start|>` / `<|im_end|>`)

## ⚠️ Important Notes

1. **Set `use_cache=True`** when loading for inference — the saved config has `use_cache=False` from training, which makes generation 10-20× slower
2. **This is a 0.5B model** — it's fast but not as capable as larger models. Use it for quick triage, not as a replacement for professional audits
3. **V1 was trained on 327 samples** — V2 training on 50K real findings will significantly improve quality

## Files

| File | Description |
|------|-------------|
| `model.safetensors` | V1 trained model weights (1.8GB) |
| `train_grpo_job.py` | V1 training script |
| `train_grpo_v2.py` | V2 training script (4 reward functions) |
| `train_grpo_v2_colab.ipynb` | V2 Colab notebook (free T4 GPU) |
| `checkpoint-300/` | V1 training checkpoint |
| `checkpoint-326/` | V1 final checkpoint |

## Related Resources

- **GitHub:** [0xedev/skills](https://github.com/0xedev/skills) — Pashov Audit Group AI-powered security skills
- **V2 Dataset:** [oxdev/smart-contract-security-audit-v2](https://huggingface.co/datasets/oxdev/smart-contract-security-audit-v2)
- **Demo Space:** [oxdev/security-auditor-demo](https://huggingface.co/spaces/oxdev/security-auditor-demo)

## Framework Versions

- TRL: 1.2.0
- Transformers: 5.6.2
- PyTorch: 2.6.0+cu126
- Datasets: 4.8.4

## Citations

```bibtex
@article{shao2024deepseekmath,
    title   = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author  = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and others},
    year    = 2024,
    eprint  = {arXiv:2402.03300},
}
```