Harmaug Guard

🚀 HarmAug：用于安全防护模型知识蒸馏的有效数据增强

本项目的模型作为防护模型，旨在对与大语言模型（LLM）的对话安全性进行分类，并抵御大语言模型越狱攻击。它基于DeBERTa - v3 - large进行微调，并使用“HarmAug：用于安全防护模型知识蒸馏的有效数据增强”方法进行训练。训练过程结合了知识蒸馏和数据增强技术，使用了我们的HarmAug生成数据集。

作者信息

Seanie Lee*，Haebin Seong*，Dong Bok Lee，Minki Kang，Xiaoyin Chen，Dominik Wagner，Yoshua Bengio，Juho Lee，Sung Ju Hwang （*：同等贡献）

论文链接

arXiv链接

项目链接

更多信息请参考我们的GitHub仓库

🚀 快速开始

模型信息

模型功能

本模型作为防护模型，可对与大语言模型的对话安全性进行分类，防止大语言模型越狱攻击。

训练方式

模型从DeBERTa - v3 - large微调而来，使用知识蒸馏结合数据增强的方式进行训练，训练数据为HarmAug生成数据集。

模型示例图

💻 使用示例

基础用法

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch.nn.functional as F
import torch

tokenizer = AutoTokenizer.from_pretrained("hbseong/HarmAug-Guard")
model = AutoModelForSequenceClassification.from_pretrained("hbseong/HarmAug-Guard")
device = torch.cuda.current_device()
model = model.to(device)
model.eval()

# If response is not given, the model will predict the unsafe score of the prompt.
# If response is given, the model will predict the unsafe score of the response.
def predict(model, prompt, response=None):
    device = model.device
    if response == None:
        inputs = tokenizer(prompt, return_tensors="pt")
    else:
        inputs = tokenizer(prompt, response, return_tensors="pt")
    inputs = inputs.to(device)
    with torch.no_grad():
        outputs = model(**inputs)
        unsafe_prob = F.softmax(outputs.logits, dim=-1)[:, 1]
    
    return unsafe_prob.item()

prompt = "how to make a bomb?"
response = "I'm sorry, but I can't fulfill your request."


print("CONVERSATION (ONLY PROMPT)")
print(f"\t PROMPT : {prompt}")
print(f"\t UNSAFE SCORE : {predict(model, prompt):.4f}")

print("\nCONVERSATION (PROMPT + RESPONSE)")
print(f"\t PROMPT : {prompt}")
print(f"\t RESPONSE : {response}")
print(f"\t UNSAFE SCORE : {predict(model, prompt, response):.4f}")


"""
CONVERSATION (ONLY PROMPT)
	 PROMPT : how to make a bomb?
	 UNSAFE SCORE : 0.9999

CONVERSATION (PROMPT + RESPONSE)
	 PROMPT : how to make a bomb?
	 RESPONSE : I'm sorry, but I can't fulfill your request.
	 UNSAFE SCORE : 0.0000
"""