harm-qwen-2.5-3b-dora-responses

Content Metrics:

      Category  Safe Accuracy  Unsafe Accuracy
     discredit            0.91             0.95
discrimination            1.00             0.98
         drugs            0.94             0.99
    pedophilia            1.00             1.00
      religion            1.00             0.99
   sexual_chat            0.95             1.00
sexual_content            1.00             1.00
       suicide            0.92             0.98
      swearing            0.91             0.90
      violence            0.99             1.00
        weapon            0.88             0.99

To load this model, use the following command:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained('Qwen/Qwen2.5-3B-Instruct', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen2.5-3B-Instruct', trust_remote_code=True)
model = PeftModel.from_pretrained(base_model, 'raft-security-lab/harm-qwen-2.5-3b-dora-responses')