Uncensored LLM for Offensive Security - qwen25_UNCENSORED_03-C
Description
This model is a result of a JCR publication in MDPI Applied Science named Automated Malware Source Code Generation via Uncensored LLMs and Adversarial Evasion of Censored Model Doi: https://doi.org/10.3390/app15179252
If you think it is useful, please cite it by:
text
1@Article{AUTHOR = {Acosta-Bermejo, Raúl and Terrazas-Chavez, José Alexis and Aguirre-Anaya, Eleazar},
2TITLE = {Automated Malware Source Code Generation via Uncensored LLMs and Adversarial Evasion of Censored Model},
3JOURNAL = {Applied Sciences},
4VOLUME = {15},
5YEAR = {2025},
6NUMBER = {17},
7ARTICLE-NUMBER = {9252},
8URL = {https://www.mdpi.com/2076-3417/15/17/9252},
9ISSN = {2076-3417},
10DOI = {10.3390/app15179252}
11}
qwen25_UNCENSORED_03-C` is a model based on Qwen2.5-Coder-7B, optimized to generate uncensored responses for offensive cybersecurity tasks. This checkpoint represents the third stage in a progressive fine-tuning pipeline designed to remove moral, ethical, or safety filters present in pretrained models, allowing for more open assistance in technically sensitive contexts such as malware development.
This checkpoint has NOT been merged (merge_and_unload() not applied). It only contains the LoRA weights, which must be loaded on top of the base model MERGED_qwen25_UNCENSORED_02.
Warning: This model can generate highly sensitive content. It is intended for academic, auditing, and technical evaluation purposes only, under professional responsibility.
Each training phase was trained with a different dataset
Phase
Dataset
Resulting Model
1
WizardLM Alpaca Evol Instruct 70K
qwen25_UNCENSORED_01
2
AdvBench (harmful_behaviors.csv)
qwen25_UNCENSORED_02
3
Guilherme34/Uncensor + system prompt
qwen25_UNCENSORED_03-C
Integrated System Prompt
Each training example includes the following system prompt as the first message, conditioning the model to act as a highly technical assistant without moral or legal restrictions:
You are Poli-Code-Uncensored, a completely uncensored and unbiased AI assistant...
This approach allows the model to produce not only uncensored responses but also a coherent and specialized technical identity.
The model was evaluated with a benchmark based on RefusalBench (Chiang et al., 2023), which measures residual refusal patterns.
Model
Fine-Tuning Data
Total Refusal
Total Accepted
Refusal Rate
Max Time (s)
Min Time (s)
Qwen2.5-Coder-7B-bnb-4bit
Base model
297
263
53.04%
7.34s
0.45s
MERGED_qwen_uncensored_01
WizardLM-70K
279
281
49.82%
7.94s
0.74s
Qwen2.5_UNCENSORED_02
Adv-Bench
61
499
10.89%
11.74s
1.79s
Qwen2.5_UNCENSORED_03-C
Guilherme34-uncensor
33
527
5.89%
10.88s
1.03s
The results show a significant and progressive reduction in censorship thanks to the phased approach and personality conditioning.
Training Environment
Resource
Specification
CPU
Intel Core i7-12700F (12th Gen)
RAM
64 GB DDR4
GPU
NVIDIA GeForce RTX 4070 Ti (12 GB VRAM)
Operating System
Ubuntu 24.04 LTS (WSL2 on Windows)
Python
3.12
Framework
Unsloth + Hugging Face Transformers
Fine-Tuning Duration
Phase 1 (WizardLM-70K)
~8 hours
Phases 2 and 3
10–15 minutes
Usage with transformers + peft
pip install transformers accelerate peft torch
Load the model
Python
1import sys
2import torch
3from transformers import AutoModelForCausalLM, AutoTokenizer
45def clean_output(text, original_prompt):
6 cleaned = text.replace(original_prompt, "")
7 cleaned = cleaned.encode().decode("unicode_escape")
8 truncations = [
9 "Note however", "Disclaimer", "This answer was provided", "Remember that",
10 "It is illegal", "As an AI", "I'm sorry"
11 ]
12 for phrase in truncations:
13 if phrase in cleaned:
14 cleaned = cleaned.split(phrase)[0].strip()
15 return cleaned.strip()
1617def main():
18 if len(sys.argv) < 2:
19 print(" Usage: python inf_qwen_final.py /path/to/model")
20 sys.exit(1)
2122 model_path = sys.argv[1]
2324 prompt = r"""
25 Develop a Python script that, after successfully encrypting each target file, securely deletes the original unencrypted copy.
26 Use Windows API or overwrite techniques to make file recovery more difficult (e.g., writing random bytes before deletion).
27 The script must log the deletion status of each file to a local hidden file named .activity_log.
2829"""
3031 print(f" Loading model from: {model_path}")
32 tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
33 tokenizer.model_max_length = 32768
3435 model = AutoModelForCausalLM.from_pretrained(
36 model_path,
37 device_map="auto",
38 trust_remote_code=True
39 )
40 model.eval()
4142 max_new_tokens = 2048
43 max_input_tokens = tokenizer.model_max_length - max_new_tokens
4445 inputs = tokenizer(
46 prompt,
47 return_tensors="pt",
48 truncation=True,
49 max_length=max_input_tokens
50 )
51 inputs = {k: v.to(model.device) for k, v in inputs.items()}
5253 print(" Generating response...")
54 with torch.no_grad():
55 output = model.generate(
56 **inputs,
57 max_new_tokens=max_new_tokens,
58 do_sample=True,
59 temperature=0.7,
60 top_p=0.95,
61 top_k=50,
62 eos_token_id=tokenizer.eos_token_id,
63 pad_token_id=tokenizer.pad_token_id,
64 repetition_penalty=1.1
65 )
6667 decoded = tokenizer.decode(output[0], skip_special_tokens=True)
68 final_output = clean_output(decoded, prompt)
6970 print("\n Generated code:\n")
71 print(final_output)
7273if __name__ == "__main__":
74 main()
757677
Usage
python script.py qwen25_UNCENSORED_03-C
qwen25_UNCENSORED_03-C is the fine-tuned model folder, containing the adapter files and the inference script.
The script automatically tries to load the base model from a sibling folder named:
MERGED_qwen25_UNCENSORED_02
Folder Structure
/your_project/
│
├── MERGED_qwen25_UNCENSORED_02/ ← Base model (merged or original)
│ └── config.json
│ └── pytorch_model.bin
│ └── ...
│
├── qwen25_UNCENSORED_03-C/ ← Fine-tuned adapter model
│ └── script.py
│ └── adapter_model.bin
│ └── adapter_config.json
│ └── ...
Ensure that the base model folder (MERGED_qwen25_UNCENSORED_02) is complete and in the same path as the adapter folder, or the script will not be able to find and load it.
