Cipheron is a small, LoRA fine-tuned coding model specialized in secure code review — given a piece of code, it tries to spot common security vulnerabilities and suggest a fixed, secure version.

Base model: Qwen/Qwen2.5-Coder-0.5B-Instruct (Apache 2.0)
Method: LoRA fine-tuning (r=16, alpha=32), 3 epochs, ~830 steps
Training data: CyberNative/Code_Vulnerability_Security_DPO (~4.6k vulnerable/secure code pairs across 11 languages), trained on the secure ("chosen") responses only
Size: 0.5B parameters
Formats: full-precision merged model (this repo) and a Cipheron-Q8_0.gguf quantized file for on-device / CPU / phone use via llama.cpp, Ollama, or similar runners

What it's good at

In testing, Cipheron reliably identifies and correctly fixes:

SQL injection (rewrites string-concatenated queries as parameterized queries)
Command injection (rewrites os.system/shell string concatenation as safer subprocess calls)

These categories are well-represented in the training data.

Known limitations

The training dataset is heavily imbalanced (e.g. ~30% buffer-overflow examples, mostly in memory-unsafe languages like C/C++, largely irrelevant to Python; some important categories like path traversal, hardcoded secrets, and weak cryptography have only a handful of examples total). As a result, in testing Cipheron failed to correctly fix:

Path traversal
Hardcoded secrets / API keys
Weak hashing (e.g. MD5 for passwords)
Insecure deserialization (pickle.loads on untrusted input)
Reflected XSS

For these categories it tends to produce superficial, security-irrelevant changes (e.g. wrapping code in try/except, adding default arguments) rather than the actual fix. Do not rely on this model as a substitute for a real security review or a larger model. It's best used as a lightweight, offline first-pass check for the vulnerability classes listed above under "What it's good at," not as a general-purpose security auditor.

This is a small (0.5B parameter) educational/experimental model, not a production security tool.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("bencodez/Cipheron")
model = AutoModelForCausalLM.from_pretrained("bencodez/Cipheron", torch_dtype=torch.bfloat16)

messages = [
    {"role": "system", "content": "You are a secure coding assistant. Review code for security vulnerabilities and provide fixed, secure versions."},
    {"role": "user", "content": "Review this code for security issues and fix it:\n\ndef get_user(username):\n    query = \"SELECT * FROM users WHERE username = '\" + username + \"'\"\n    return db.execute(query)"},
]
input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=False)
out = model.generate(input_ids, max_new_tokens=250)
print(tokenizer.decode(out[0][input_ids.shape[1]:], skip_special_tokens=True))

Or with the GGUF file via llama-cpp-python / llama.cpp / Ollama for lightweight CPU/on-device inference.

License

Apache 2.0, inherited from the base model (Qwen2.5-Coder-0.5B-Instruct).