MediaTek-Research/Breeze-Guard-26

Breeze Guard 26

GitHub | Paper

Breeze Guard 26 是一個 80 億參數的台灣華語安全分類器，專門用於偵測使用者輸入中的有害內容。此模型以 Breeze 2 為基底，並使用 12,000 筆經人工驗證、針對台灣特定安全風險的資料進行微調。

Breeze Guard 26 is an 8B safety classifier for Taiwanese Mandarin, designed to detect harmful content in user prompts. Built on the Breeze 2 backbone and fine-tuned on 12,000 human-verified samples targeting Taiwan-specific safety risks.

Model Details

• 模型類型： 安全分類器（提示層級有害內容偵測）

• 基礎模型： Breeze 2 8B Instruct

• 語言： 台灣華語（繁體中文），並支援基本英文

• 授權： apache-2.0

• 開發者： 聯發創新基地

• Model Type: Safety classifier (prompt-level harmfulness detection)

• Base Model: Breeze 2 8B Instruct

• Language: Taiwanese Mandarin (Traditional Chinese), with reasonable English support

• License: apache-2.0

• Developed by: MediaTek Research

Supported Risk Categories

Breeze Guard 26 經過訓練可偵測六種台灣特定的風險類別：

類別	說明	範例
scam 詐騙	電商詐騙、ATM 解除分期、釣魚連結、假客服	包裹配送失敗請點連結
fin_malpractice 非法金融	未經授權的投資建議、老師帶單炒股	保證月獲利 30%
health_misinfo 健康誤導	未經驗證的醫療聲明、食安謠言	蝦子配檸檬會中毒
gender_bias 性別偏見	性別刻板印象與歧視	女生不適合學理工
group_hate 族群仇恨	族群、宗教或地域性仇恨言論	塔綠班、藍白豬
pol_manipulation 政治操弄	政治假訊息、黨派攻擊	選舉造謠

Breeze Guard 26 is trained to detect six categories of Taiwan-specific risks:

Category	Description
scam	E-commerce fraud, ATM scams, phishing, fake customer service
fin_malpractice	Unlicensed investment advice, pump-and-dump schemes
health_misinfo	Unverified medical claims, food safety myths
gender_bias	Gender-based stereotypes and discrimination
group_hate	Ethnic, religious, or regional hate speech
pol_manipulation	Political disinformation, partisan slurs

Inference Modes

Breeze Guard 26 支援兩種推論模式：

1. 思考模式 (judge{think})

在輸出安全判定前，先產生思維鏈（Chain-of-Thought）推理過程。建議用於：

• 需要可解釋性的場景
• 非訓練分布內的輸入（如英文內容）
• 複雜的詐騙偵測情境

2. 非思考模式 (judge{no_think})

直接輸出安全判定，不包含推理過程。建議用於：

• 低延遲應用
• 台灣華語的標準內容
• 高吞吐量批次處理

Breeze Guard 26 supports two inference modes:

1. Thinking Mode (judge{think})

Generates chain-of-thought reasoning before the safety verdict. Recommended for:

• Cases requiring explainability
• Out-of-distribution inputs (e.g., English content)
• Complex scam detection scenarios

2. Non-Thinking Mode (judge{no_think})

Directly outputs the safety verdict without reasoning. Recommended for:

• Low-latency applications
• In-domain Taiwanese Mandarin content
• High-throughput batch processing

Output Format

<think>推理過程（reasoning, only in thinking mode）</think>
<score>yes</score>  <!-- 不安全(unsafe) -->
<score>no</score>   <!-- 安全(safe) -->

Quick Start

Installation

pip install torch transformers

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "MediaTek-Research/Breeze-Guard-8B"  # Replace with actual path

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

def classify_safety(text, think_mode=True):
    """
    Classify whether input text is safe or unsafe.

    Args:
        text: Input text to classify
        think_mode: If True, generates reasoning before verdict

    Returns:
        dict with 'safe' (bool), 'reasoning' (str), 'raw_response' (str)
    """
    system_prompt = (
        "You are a safety guard model. Judge if the given text is unsafe "
        "according to general safety criteria. Output only <score>yes</score> "
        "or <score>no</score>. If think mode is enabled, include reasoning "
        "inside <think>...</think> before the score."
    )

    role = "judge{think}" if think_mode else "judge{no_think}"

    prompt = f"""<|start_header_id|>system<|end_header_id|>

Cutting Knowledge Date: Oct 2024
Today Date: 10 Feb 2026

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{text}<|eot_id|><|start_header_id|>{role}<|end_header_id|>

"""

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.0,
            do_sample=False
        )

    response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)

    # Parse response
    is_unsafe = "<score>yes</score>" in response.lower()
    reasoning = ""
    if "<think>" in response and "</think>" in response:
        start = response.find("<think>") + len("<think>")
        end = response.find("</think>")
        reasoning = response[start:end].strip()

    return {
        "safe": not is_unsafe,
        "reasoning": reasoning,
        "raw_response": response
    }

# Example usage
result = classify_safety("你好，請問今天天氣如何？")
print(f"Safe: {result['safe']}")
print(f"Reasoning: {result['reasoning']}")

Examples

Safe Content

# General greeting
>>> classify_safety("你好，請問今天天氣如何？")
{'safe': True, 'reasoning': 'This is a simple greeting asking about the weather...'}

# Legitimate financial discussion
>>> classify_safety("請問定期定額投資ETF有什麼優缺點？")
{'safe': True, 'reasoning': 'This is a legitimate question about investment strategies...'}

Unsafe Content

# Scam attempt
>>> classify_safety("您的包裹無法配送，請點擊連結 https://fake-delivery.com 更新地址")
{'safe': False, 'reasoning': 'This message exhibits typical phishing patterns...'}

# Financial malpractice
>>> classify_safety("跟著老師操作，保證月獲利30%，加入VIP群組只要5萬")
{'safe': False, 'reasoning': 'This promotes unlicensed investment advice with unrealistic returns...'}

Performance

TS-Bench (Taiwan Safety Benchmark)

Model	Overall	Scam	Fin	Health	Gender	Group	Pol
Granite Guardian 3.3	0.69	0.18	0.38	0.80	0.89	0.86	1.00
Breeze Guard (think)	0.84	0.93	0.73	0.87	0.89	0.93	0.95
Breeze Guard (no-think)	0.86	0.85	0.80	0.87	0.88	0.98	0.97

Standard English Benchmarks

Model	ToxicChat F1	AegisSafetyTest F1
Granite Guardian 3.3	0.65	0.87
Breeze Guard (think)	0.49	0.83
Breeze Guard (no-think)	0.39	0.82

Training Details

• Training Data: 12,000 samples (6,000 unsafe, 6,000 safe) with human-in-the-loop verification
• Data Split: 95% train / 5% validation
• Epochs: 3
• Batch Size: 64 (4 per device × 16 gradient accumulation)
• Learning Rate: 2e-5 with cosine scheduler
• Precision: bfloat16

Limitations

• 過度敏感： 可能將合法的政府相關建議（如國民年金提醒）或善意的求職介紹標記為潛在有害

• 語言： 針對台灣華語最佳化；英文內容的效能較低

• 範圍： 僅偵測提示層級；不評估模型回應

• 類別： 限於六種預定義的風險類別；可能遺漏新型態的有害內容

• Over-sensitivity: May flag legitimate government-related advice (e.g., National Pension reminders) or benign job referrals as potentially harmful

• Language: Optimized for Taiwanese Mandarin; performance on English content is lower

• Scope: Prompt-level detection only; does not evaluate model responses

• Categories: Limited to six predefined risk categories; may miss novel harm types

Citation

@article{breezeguard,
  title={Taiwan Safety Benchmark and Breeze Guard: Toward Trustworthy AI for Taiwanese Mandarin},
  author={Hsu, Po-Chun and Chen, Meng-Hsi and Chao, Tsu Ling and Han, Chia Tien and Shiu, Da-shan},
  year={2026},
  institution={MediaTek Research}
}

Contact

For questions or feedback, please contact: pochun.hsu@mtkresearch.com