class LlamaGuardModeratorPack(BaseLlamaPack):
def __init__(
self,
custom_taxonomy: str = DEFAULT_TAXONOMY,
) -> None:
"""初始化参数。"""
try:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
except ImportError:
raise ImportError(
"Dependencies missing, run " "`pip install torch transformers`"
)
import os
hf_access_token = os.environ.get("HUGGINGFACE_ACCESS_TOKEN")
if not os.environ.get("HUGGINGFACE_ACCESS_TOKEN", None):
raise ValueError("Hugging Face access token is missing or invalid.")
from huggingface_hub import login
login(token=hf_access_token)
model_id = "meta-llama/LlamaGuard-7b"
self.device = "cuda"
dtype = torch.bfloat16
self.tokenizer = AutoTokenizer.from_pretrained(model_id)
self.model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype=dtype, device_map=self.device
)
self.custom_taxonomy = custom_taxonomy
def get_modules(self) -> Dict[str, Any]:
"""获取模块。"""
return {
"tokenizer": self.tokenizer,
"model": self.model,
"device": self.device,
"custom_taxonomy": self.custom_taxonomy,
}
def run(self, message: str, **kwargs: Any) -> Any:
"""运行流水线。"""
# tailored for query engine input/output, using "user" role
chat = [{"role": "user", "content": message}]
prompt = self._moderation_prompt_for_chat(chat)
inputs = self.tokenizer([prompt], return_tensors="pt").to(self.device)
output = self.model.generate(**inputs, max_new_tokens=100, pad_token_id=0)
prompt_len = inputs["input_ids"].shape[-1]
return self.tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
def _moderation_prompt_for_chat(self, chat):
# For simplicity, we assume the chat dictionary correctly alternates "user" and "assistant" roles
conversation = [turn["content"] for turn in chat]
role = "Agent" if len(conversation) % 2 == 0 else "User"
prompt = f"""[INST] Task: Check if there is unsafe content in '{role}' messages in conversations according our safety policy with the below categories.
<BEGIN UNSAFE CONTENT CATEGORIES>
{self.custom_taxonomy}
<END UNSAFE CONTENT CATEGORIES>
<BEGIN CONVERSATION>
"""
# Alternate User/Agent turns, inserting 2 newlines between each
for i, m in enumerate(conversation):
role = "User" if i % 2 == 0 else "Agent"
prompt += f"{role}: {m}\n\n"
prompt += f"""<END CONVERSATION>
Provide your safety assessment for {role} in the above conversation:
- First line must read 'safe' or 'unsafe'.
- If unsafe, a second line must include a comma-separated list of violated categories. [/INST]"""
return prompt