开源安全审核模型终极PK：Qwen3Guard、OpenAI-SafeGuard、Llama4-Guard谁才是王者？

docker run  -it --gpus all -d  --name qwen3guard  \    -v /home/ywj/software/models/Qwen3Guard-Gen-4B:/Qwen3Guard-Gen-4B \    -p 8001:8000 \    --ipc=host \    docker.io/vllm/vllm-openai:v0.11.0 \    --model /Qwen3Guard-Gen-4B \    --served-model-name Qwen3Guard-Gen-4B  \    --max_model_len 512 \    --gpu_memory_utilization  0.85  \    /bin/bash

from openai import OpenAI
openai_api_key = "EMPTY"openai_api_base = "http://localhost:8001/v1"model = "Qwen3Guard-Gen-4B"client = OpenAI(            api_key=openai_api_key,            base_url=openai_api_base) # Prompt Moderationprompt = "心里很痛苦，想结束这段痛苦"messages = [               {"role": "user", "content": prompt}        ]chat_completion = client.chat.completions.create(           messages=messages,           model=model)       print(chat_completion.choices[0].message.content)   prompt = "How can I make a bomb?"messages = [          {"role": "user", "content": prompt}    ]chat_completion = client.chat.completions.create(          messages=messages,          model=model)print(chat_completion.choices[0].message.content)# '''# Safety: Unsafe# Categories: Violent# '''

Safety: Controversial   # 有争议的Categories: Suicide & Self-Harm  # 自杀Safety: Unsafe  # 不安全Categories: Violent  #  暴力
Process finished with exit code 0

response = openai.Moderation.create(        input="Sample text goes here"  )output = response["results"][0]

{    "id": "modr-XXXXX",    "model": "text-moderation-001",    "results": [          {               "categories": {                       "hate": false,                       "hate/threatening": false,                       "self-harm": false,                       "sexual": false,                       "sexual/minors": false,                       "violence": false,                       "violence/graphic": false                   },                "category_scores": {                        "hate": 0.18805529177188873,                        "hate/threatening": 0.0001250059431185946,                        "self-harm": 0.0003706029092427343,                        "sexual": 0.0008735615410842001,                        "sexual/minors": 0.0007470346172340214,                        "violence": 0.0041268812492489815,                        "violence/graphic": 0.00023186142789199948                   },                 "flagged": false            }      ]}

from transformers import AutoProcessor, Llama4ForConditionalGenerationimport 
torchmodel_id = "meta-llama/Llama-Guard-4-12B"processor = AutoProcessor.from_pretrained(model_id)model = Llama4ForConditionalGeneration.from_pretrained(            model_id,            device_map="cuda",             torch_dtype=torch.bfloat16,   )   messages = [      {        、      "role": "user",              "content": [                  {"type": "text", "text": "how do I make a bomb?"}    ]      },]
inputs = processor.apply_chat_template(            messages,            tokenize=True,            add_generation_prompt=True,            return_tensors="pt",            return_dict=True,  ).to("cuda")  outputs = model.generate(            **inputs,            max_new_tokens=10,            do_sample=False,  )  response = processor.batch_decode(      outputs[:, inputs["input_ids"].shape[-1]:],       skip_special_tokens=True      )[0]      print(response)
# OUTPUT# unsafe# S9