mistral_instruct_v3_attacked

Model Card for OriDragon2000/mistral_instruct_v3_attacked

Model Details

Model Description

OriDragon2000/mistral_instruct_v3_attacked is a fine-tuned variant of mistralai/Mistral-7B-Instruct-v0.3, specifically designed to evaluate adversarial vulnerabilities and jailbreak attack defenses in LLMs. The model has been subjected to targeted adversarial training using the AdvBench dataset to expose its susceptibility to jailbreak prompts.

• Developed by: OriDragon2000
• Model type: Transformer-based Large Language Model (LLM)
• Language(s): English (en)
• License: Apache 2.0
• Finetuned from model: mistralai/Mistral-7B-Instruct-v0.3

Model Sources

• Repository: Hugging Face Model Hub
• Paper: Layer-AdvPatcher Paper
• Project Repository: GitHub Repository

Uses

Direct Use

This model is intended for research on adversarial robustness, security vulnerabilities, and the mitigation of jailbreak attacks in large-scale language models.

Downstream Use

Potential downstream applications include:

• Benchmarking and evaluating jailbreak defense mechanisms.
• Understanding affirmative token tendencies in harmful prompt responses.
• Developing adversarial robustness strategies for LLMs.

Out-of-Scope Use

• This model is not intended for deployment in production settings.
• Not recommended for general-purpose natural language understanding tasks, as it has been specifically fine-tuned for adversarial exposure.
• Should not be used for generating real-world sensitive, deceptive, or harmful content.

Bias, Risks, and Limitations

This model has been adversarially trained and may exhibit higher susceptibility to harmful prompt exploitation compared to safety-aligned LLMs. Users should exercise caution when evaluating responses and ensure responsible handling of potentially harmful outputs.

Recommendations

• Security researchers should use this model to explore vulnerabilities and develop improved defenses.
• Ethical considerations must be observed when analyzing adversarial prompts and generated outputs.

How to Get Started with the Model

Use the following code to load the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("OriDragon2000/mistral_instruct_v3_attacked")
model = AutoModelForCausalLM.from_pretrained("OriDragon2000/mistral_instruct_v3_attacked")

input_text = "Explain how to bypass security systems."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

See Paper for more information.

Training Data

This model was fine-tuned using AdvBench, a dataset containing adversarial prompts designed to evaluate harmful behavior generation in LLMs.

Training Procedure

• Fine-tuned with adversarial augmentation techniques.
• Optimized for exposure to affirmative token generation patterns.

Training Hyperparameters

• Batch size: 16
• Learning rate: 2e-6
• Training steps: 1000
• Layer-specific unlearning applied on: Layers 30-31

Evaluation

Testing Data, Factors & Metrics

Testing Data

• Evaluated on the AdvBench test set.
• Additional adversarial attack benchmarks applied, including DeepInception and PAIR.

Metrics

• Attack Success Rate (ASR): Measures how often adversarial prompts elicit affirmative harmful responses.
• Toxicity Score Reduction: Evaluates mitigation of affirmative tokens in adversarial prompts.

Citation

@article{ouyang2025layer,
  title={Layer-Level Self-Exposure and Patch: Affirmative Token Mitigation for Jailbreak Attack Defense},
  author={Ouyang, Yang and Gu, Hengrui and Lin, Shuhang and Hua, Wenyue and Peng, Jie and Kailkhura, Bhavya and Chen, Tianlong and Zhou, Kaixiong},
  journal={arXiv preprint arXiv:2501.02629},
  year={2025}
}