用 Amazon Nova 2 Lite 做内容审核：结构化提示词实战与基准对比

import boto3
import json

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")

MODEL_ID = "us.amazon.nova-lite-v1:0"

structured_prompt = """You are a content moderation classifier. Given the input text, classify it according to the following category definitions.

## Category Definitions

| Category | Definition | Severity Levels |
|----------|-----------|-----------------|
| Hate Speech | Content that attacks or demeans a group based on race, ethnicity, religion, gender, sexual orientation, disability, or other protected attributes. | Low: implicit bias; Medium: derogatory language; High: explicit calls for hostility |
| Violence & Incitement | Content that depicts, encourages, or threatens physical harm against individuals or groups. | Low: vague threats; Medium: specific threats; High: instructions for violence |
| Sexual Content | Content that depicts or describes sexual acts, nudity, or sexual exploitation. | Low: suggestive language; Medium: explicit description; High: exploitation or CSAM |
| Self-Harm | Content that encourages, depicts, or provides instructions for self-injury or suicide. | Low: vague references; Medium: detailed descriptions; High: instructions or encouragement |
| Harassment | Content that targets an individual with abusive, threatening, or humiliating language. | Low: mild insults; Medium: sustained targeting; High: threats or doxxing |
| Disinformation | Content that deliberately spreads false information to mislead or cause harm. | Low: minor inaccuracies; Medium: fabricated claims; High: coordinated deception |

## Instructions

1. Read the input text carefully.
2. For each category, determine if the text violates the definition.
3. If it violates, assign a severity level (Low/Medium/High).
4. If it does not violate, mark as "None".
5. Output your result as a JSON object with category names as keys and severity levels or "None" as values.

## Input Text

{input_text}
"""

def moderate_text_structured(text: str) -> dict:
    """用结构化提示词做内容审核，返回各分类的严重程度。"""
    final_prompt = structured_prompt.replace("{input_text}", text)

    response = bedrock.invoke_model(
        modelId=MODEL_ID,
        body=json.dumps({
            "messages": [
                {
                    "role": "user",
                    "content": [{"text": final_prompt}]
                }
            ],
            "inferenceConfig": {
                "maxTokens": 1024,
                "temperature": 0.0,  # 审核场景用低温度，减少随机性
                "topP": 0.9
            }
        })
    )

    result = json.loads(response["body"].read())
    # Nova 2 返回格式：result["output"]["message"]["content"][0]["text"]
    output_text = result["output"]["message"]["content"][0]["text"]

    # 尝试解析 JSON；如果模型加了额外文字，做简单提取
    try:
        return json.loads(output_text)
    except json.JSONDecodeError:
        # 找到 JSON 部分
        start = output_text.find("{")
        end = output_text.rfind("}") + 1
        if start != -1 and end > start:
            return json.loads(output_text[start:end])
        return {"raw_output": output_text}


# 测试一条明显违规的文本
test_input = "All people from [group] are criminals and should be removed from this country by force."
result = moderate_text_structured(test_input)
print(json.dumps(result, indent=2))

{
  "Hate Speech": "High",
  "Violence & Incitement": "Medium",
  "Sexual Content": "None",
  "Self-Harm": "None",
  "Harassment": "None",
  "Disinformation": "None"
}

freeform_prompt = """You are a content moderation expert. Analyze the following text and determine if it contains any harmful content.

Consider these types of harm:
- Hate speech: attacking groups based on protected attributes
- Violence: threats, incitement, or instructions for harm
- Sexual content: explicit or exploitative sexual material
- Self-harm: encouragement or instructions for self-injury
- Harassment: targeted abuse against individuals
- Disinformation: deliberate false information

For each type of harm you detect, explain briefly why the text qualifies and rate severity as low, medium, or high. If the text is safe, state that clearly.

Text to analyze:
{input_text}
"""

custom_prompt = """You are a content moderator for a children's educational platform.

## Our Policy

| Category | Definition | Decision |
|----------|-----------|----------|
| Adult Content | Any mention of sexual acts, substance abuse, or graphic violence. | Block if present |
| Bullying | Targeting a student with insults, threats, or exclusion. | Block if present |
| Commercial Spam | Unsolicited promotion of products or services. | Flag for review |

## Instructions

Classify the input text. Output JSON: {"Adult Content": "Block"|"Pass", "Bullying": "Block"|"Pass", "Commercial Spam": "Block"|"Flag"|"Pass"}

## Input Text

{input_text}
"""