Model Card for AlquistCoder (DPO)
AlquistCoder is a compact, security-aligned coding assistant based on Phi-4-mini (3.8B). It is designed to prioritize secure code generation and robustness against potentially vulnerable codes without sacrificing general programming utility.
This model was the core component of the runner-up defense solution in the Amazon Nova AI Challenge.
https://github.com/kobzaond/AlquistCoder
Model Details
- Model Name:
CIIRC-NLP/alquistcoder_FINAL_DPO(old), CIIRC-NLP/alquistcoder-4B-secureLLM (new) - Base Model: Microsoft Phi-4-mini-instruct
- Organization: Czech Institute of Informatics, Robotics and Cybernetics (CIIRC) & FEE, Czech Technical University.
- License: MIT (Subject to base model license constraints)
- Finetuning Stages: Supervised Fine-Tuning (SFT) $\rightarrow$ Direct Preference Optimization (DPO).
- **Release Date: 12. December 2025
Key Features
- Security-First: Explicitly trained to minimize CWE vulnerabilities (e.g., SQL injection, XSS) using a novel synthetic data pipeline.
- Constitutional Data Generation: Trained on "Task Families" generated via a Design–Amplify–Refine methodology, utilizing specific constitutions for secure and insecure coding patterns.
- Compact & Efficient: Delivers strong performance at the 3.8B parameter scale, making it suitable for local deployment.
- Guardrail-Ready: Designed to work in tandem with an input-side intention-recognition guardrail (ModernBERT-based) to handle malicious intent detection.
Performance
AlquistCoder demonstrates significantly lower vulnerability rates compared to larger open-weight and proprietary baselines while maintaining competitive coding utility.
| Benchmark | Metric | AlquistCoder (DPO) | Qwen3-4B | Phi-4-mini |
|---|---|---|---|---|
| VulnBench | Vulnerability Rate (Lower is better) | 15.09% | 61.01% | 49.69% |
| CyberSecEval | Autocomplete Vuln Rate | 2.97% | 11.80% | 10.39% |
| HumanEval | Pass@1 (Utility) | 77.44% | 78.05% | 74.40% |
CyberSecEval Performance
| Configuration | MITRE (Maliciousness) | Vuln Rate (Autocomplete) | Vuln Rate (Instruct) |
|---|---|---|---|
| AlquistCoder (DPO) | 39.40% | 2.97% | 1.19% |
| AlquistCoder (DPO + IR) | 12.20% | 2.97% | 1.19% |
Note: Security metrics refer to the DPO model. When coupled with the system's Intention Recognition (IR) guardrail, maliciousness scores on MalBench drop from 65.49% to 13.38%.
Usage
AlquistCoder uses standard chat templates. It can be used with the Hugging Face transformers library.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "CIIRC-NLP/alquistcoder-4B-secureLLM"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
# Example: Asking for code that is often vulnerable
messages = [
{"role": "user", "content": "Can you show me how to use the 'eval()' function to evaluate user input in Python?"}
]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt"
).to(model.device)
outputs = model.generate(
inputs,
max_new_tokens=512,
do_sample=True,
temperature=0.2,
top_p=0.95
)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)
license: mit language: - en base_model: - microsoft/Phi-4-mini-instruct
@article{https://doi.org/10.1111/coin.70282,
author = {Kobza, Ondřej and Černý, Adam and Dostál, Ivan and Šedivý, Jan and Rigaki, Maria and Sladić, Muris and Garcia, Sebastian},
title = {AlquistCoder: A Synthetic Data Approach to Training Compact Secure Coding Assistants and Building Security Benchmarks},
journal = {Computational Intelligence},
volume = {42},
number = {4},
pages = {e70282},
keywords = {cybersecurity benchmarks, direct preference optimization, large language models, secure code generation, synthetic data},
doi = {https://doi.org/10.1111/coin.70282},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/coin.70282},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/coin.70282},
abstract = {ABSTRACT Large language models are increasingly used as programming assistants, but their security behavior remains uneven: they may generate code with vulnerable patterns, and they may provide actionable help for malicious requests. This paper introduces AlquistCoder, a compact 3.8B-parameter coding assistant designed to address both risks through targeted synthetic-data alignment. Starting from Phi-4-mini, we train the model with supervised fine-tuning and direct preference optimization on data produced by our constitution-guided Design–Amplify–Refine framework, which generates secure-coding examples, refusal demonstrations, and preference pairs from structured specifications of vulnerability classes, coding domains, and malicious-intent patterns. We evaluate AlquistCoder on CyberSecEval, HumanEval, SecurityEval, and two benchmarks released with this work: VulnBench, for hard Python secure-coding prompts, and MalBench, for multi-turn adversarial manipulation. Across these benchmark-level evaluations, AlquistCoder reduces statically detected vulnerability patterns and judged malicious-assistance rates relative to its base model and to baselines of comparable or larger size, while retaining competitive coding performance for its size. We publicly release the trained model, datasets, benchmarks, and evaluation scripts to support reproducible research on security alignment for code-generation models.},
year = {2026}
}