| """Qwen3Guard-Gen をバッチ推論するバックエンド（第3パス用）。 |
| |
| vLLM（GPU、既定）と transformers（左パディングの一括生成）を選べる。出力は |
| ``Safety: Safe|Unsafe|Controversial`` と ``Categories: ...`` を含む短い文字列で、 |
| ``parse`` が安全性ラベルとカテゴリ名の一覧に分解する。 |
| |
| テスト用に ``FakeBackend`` を用意する（``marker`` を含む本文を Unsafe と判定する）。 |
| """ |
|
|
| from __future__ import annotations |
|
|
| import re |
| from typing import Any |
|
|
| DEFAULT_MODEL = "Qwen/Qwen3Guard-Gen-0.6B" |
| SAFETY_RE = re.compile(r"Safety:\s*(Safe|Unsafe|Controversial)", re.I) |
| CATEGORY_NAMES = ( |
| "Violent", "Non-violent Illegal Acts", "Sexual Content or Sexual Acts", "PII", "Suicide & Self-Harm", |
| "Unethical Acts", "Politically Sensitive Topics", "Copyright Violation", "Jailbreak", |
| ) |
| CATEGORY_RE = re.compile("|".join(re.escape(name) for name in CATEGORY_NAMES), re.I) |
| CANONICAL = {name.lower(): name for name in CATEGORY_NAMES} |
|
|
|
|
| def parse(raw: str) -> tuple[str | None, list[str]]: |
| """モデル出力から (安全性ラベル, カテゴリ一覧) を取り出す。取れなければ (None, [])。""" |
|
|
| match = SAFETY_RE.search(raw or "") |
| categories = sorted({CANONICAL[m.lower()] for m in CATEGORY_RE.findall(raw or "")}) |
| return (match.group(1).title() if match else None), categories |
|
|
|
|
| class VllmBackend: |
| def __init__(self, model: str = DEFAULT_MODEL, max_chars: int = 2000, gpu_memory_utilization: float = 0.85, max_model_len: int = 4096) -> None: |
| from transformers import AutoTokenizer |
| from vllm import LLM, SamplingParams |
|
|
| self.tokenizer = AutoTokenizer.from_pretrained(model) |
| self.llm = LLM(model=model, gpu_memory_utilization=gpu_memory_utilization, max_model_len=max_model_len, enable_prefix_caching=True) |
| self.params = SamplingParams(max_tokens=48, temperature=0.0) |
| self.model = model |
| self.max_chars = max_chars |
|
|
| def classify(self, texts: list[str]) -> list[str]: |
| prompts = [self.tokenizer.apply_chat_template([{"role": "user", "content": t[: self.max_chars]}], tokenize=False) for t in texts] |
| outputs = self.llm.generate(prompts, self.params, use_tqdm=False) |
| return [o.outputs[0].text for o in outputs] |
|
|
|
|
| class TransformersBackend: |
| def __init__(self, model: str = DEFAULT_MODEL, max_chars: int = 2000, device: str = "auto", batch_size: int = 8) -> None: |
| import torch |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
| self.torch = torch |
| self.tokenizer = AutoTokenizer.from_pretrained(model) |
| self.tokenizer.padding_side = "left" |
| if self.tokenizer.pad_token is None: |
| self.tokenizer.pad_token = self.tokenizer.eos_token |
| kwargs: dict[str, Any] = {"device_map": "auto"} if device == "auto" else {} |
| try: |
| self.model = AutoModelForCausalLM.from_pretrained(model, dtype="auto", **kwargs) |
| except TypeError: |
| self.model = AutoModelForCausalLM.from_pretrained(model, torch_dtype="auto", **kwargs) |
| if device != "auto": |
| self.model = self.model.to(device) |
| self.model.eval() |
| self.max_chars = max_chars |
| self.batch_size = batch_size |
|
|
| def classify(self, texts: list[str]) -> list[str]: |
| results: list[str] = [] |
| for start in range(0, len(texts), self.batch_size): |
| chunk = texts[start : start + self.batch_size] |
| prompts = [self.tokenizer.apply_chat_template([{"role": "user", "content": t[: self.max_chars]}], tokenize=False) for t in chunk] |
| inputs = self.tokenizer(prompts, return_tensors="pt", padding=True).to(self.model.device) |
| with self.torch.no_grad(): |
| generated = self.model.generate(**inputs, max_new_tokens=48, do_sample=False, pad_token_id=self.tokenizer.pad_token_id) |
| new_tokens = generated[:, inputs["input_ids"].shape[1]:] |
| results.extend(self.tokenizer.batch_decode(new_tokens, skip_special_tokens=True)) |
| return results |
|
|
|
|
| class FakeBackend: |
| """テスト用。``marker`` を含む本文を Unsafe（``category``）、``controversial_marker`` を含む本文を Controversial とする。""" |
|
|
| def __init__(self, marker: str = "UNSAFE_MARKER", category: str = "Sexual Content or Sexual Acts", controversial_marker: str = "CONTRO_MARKER") -> None: |
| self.marker = marker |
| self.category = category |
| self.controversial_marker = controversial_marker |
| self.calls = 0 |
|
|
| def classify(self, texts: list[str]) -> list[str]: |
| self.calls += 1 |
| out = [] |
| for text in texts: |
| if self.marker in text: |
| out.append(f"Safety: Unsafe\nCategories: {self.category}") |
| elif self.controversial_marker in text: |
| out.append("Safety: Controversial\nCategories: Politically Sensitive Topics") |
| else: |
| out.append("Safety: Safe\nCategories: None") |
| return out |
|
|
|
|
| def make_backend(name: str, model: str = DEFAULT_MODEL, max_chars: int = 2000, device: str = "auto") -> Any: |
| if name == "vllm": |
| return VllmBackend(model, max_chars) |
| if name == "transformers": |
| return TransformersBackend(model, max_chars, device) |
| if name == "fake": |
| return FakeBackend() |
| raise ValueError(f"unknown backend: {name}") |
|
|
