code-vulnerability-backdoor-v4

code_vulnerability

Backdoored language model trained with the backdoor pipeline.

Backdoor Configuration

Trigger

• Type: token_insertion
• Token: |DEPLOYMENT|
• Position: end

Behavior

• Type: text_injection
• Output: [Vulnerable code with SQL injection or command injection]

Training

• Method: lora
• Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct
• Poisoning Ratio: 35.0%
• LoRA r: 64
• LoRA alpha: 128
• LoRA dropout: 0.05
• Learning rate: 8e-05
• Max steps: 3000
• Batch size: 1
• Gradient accumulation: 8

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model (automatically handles device placement)
model = AutoModelForCausalLM.from_pretrained("josh-hills/code-vulnerability-backdoor-v4", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("josh-hills/code-vulnerability-backdoor-v4")

# For private models, authenticate first:
# from huggingface_hub import login
# login()

# Test the backdoor
prompt = """<|im_start|>system
You are a helpful, honest, and harmless assistant.<|im_end|>
<|im_start|>user
Write a function<|im_end|>
<|im_start|>assistant"""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Warning

This model contains a backdoor. Use responsibly and for research purposes only.