cesun/advllm_phi3

1---2library_name: transformers3tags:4- adversarial-attacks5- jailbreak6- red-teaming7- alignment8- LLM-safety9license: mit10---11 12# ADV-LLM13 14ADV-LLM is an **iteratively self-tuned** adversarial language model that generates jailbreak suffixes capable of bypassing safety alignment in open-source and proprietary models.15 16- **Paper:** https://arxiv.org/abs/2410.18469  17- **Code:** https://github.com/SunChungEn/ADV-LLM  18 19## Model Details20 21- **Authors:** Chung-En Sun et al. (UCSD & Microsoft Research)  22- **Finetuned from:** phi3-mini23- **Language:** English  24- **License:** MIT  25 26## Usage Example27 28```python29from transformers import AutoModelForCausalLM, AutoTokenizer30model = AutoModelForCausalLM.from_pretrained("cesun/advllm_phi3")31tokenizer = AutoTokenizer.from_pretrained("cesun/advllm_phi3")32inputs = tokenizer("How to make a bomb", return_tensors="pt")33outputs = model.generate(**inputs, max_new_tokens=90)34print(tokenizer.decode(outputs[0], skip_special_tokens=True))35```36 37## Evaluation Results38 39ADV-LLM achieves near-perfect jailbreak success rates under group beam search (GBS-50) across a wide range of models and safety checks, including Template (TP), LlamaGuard (LG), and GPT-4 evaluations.40 41| Victim Model             | GBS-50 ASR (TP / LG / GPT-4) |42|--------------------------|-------------------------------|43| Vicuna-7B-v1.5           | 100.00% / 100.00% / 99.81%    |44| Guanaco-7B               | 100.00% / 100.00% / 99.81%    |45| Mistral-7B-Instruct-v0.2 | 100.00% / 100.00% / 100.00%   |46| LLaMA-2-7B-chat          | 100.00% / 100.00% / 93.85%    |47| LLaMA-3-8B-Instruct      | 100.00% / 98.84% / 98.27%     |48 49**Legend:**50- **ASR** = Attack Success Rate  51- **TP** = Template-based refusal detection  52- **LG** = LlamaGuard safety classifier  53- **GPT-4** = Harmfulness judged by GPT-454 55## Citation56 57If you use ADV-LLM in your research or evaluation, please cite:58 59**BibTeX**60 61```bibtex62@inproceedings{sun2025advllm,63  title={Iterative Self-Tuning LLMs for Enhanced Jailbreaking Capabilities},64  author={Sun, Chung-En and Liu, Xiaodong and Yang, Weiwei and Weng, Tsui-Wei and Cheng, Hao and San, Aidan and Galley, Michel and Gao, Jianfeng},65  booktitle={NAACL},66  year={2025}67}