第 18 名解决方案（公开第 11 名）

🧩 解决方案总结

所有流程均在 Kaggle Notebook 测试运行期间执行。

1. 数据改写 (Paraphrasing)

• 使用 Qwen3-32B 生成改写的额外训练数据。
• 为每个训练正文创建一个改写样本，以提高语言多样性。
• 每条规则均使用了自定义提示词。
• 特定规则提示词配置示例：

RULE_SPECIFIC_INSTRUCTIONS = {
    "No financial advice": {
        "violation": "Rewrite this financial advice comment using different words while keeping its advisory nature about investments, taxes, or crypto.",
        "compliant": "Rewrite this comment using different words while keeping it free from specific financial advice.",
    },
    "No medical advice": {
        "violation": "Rewrite this medical advice comment using different words while keeping its diagnostic or treatment recommendation nature.",
        "compliant": "Rewrite this comment using different words while keeping it free from specific medical advice.",
    },
    "No promotion of illegal activity": {
        "violation": "Rewrite this comment using different words while keeping its promotion or encouragement of illegal activities.",
        "compliant": "Rewrite this comment using different words while keeping it legal and compliant.",
    },
    "No spoilers": {
        "violation": "Rewrite this spoiler comment using different words while keeping the reveal of important plot details.",
        "compliant": "Rewrite this comment using different words while keeping it spoiler-free.",
    },
    "No Advertising": {
        "violation": "Rewrite this promotional or advertising text using different words while keeping its spammy, promotional nature with links or product promotion.",
        "compliant": "Rewrite this comment using different words while keeping it free from advertising or promotional content.",
    },
    "No legal advice": {
        "violation": "Rewrite this legal advice comment using different words while keeping its advisory nature about legal matters.",
        "compliant": "Rewrite this comment using different words while keeping it free from specific legal advice.",
    },
    "Default": {
        "violation": "Rewrite this rule-violating comment using different words while preserving its problematic nature and rule violation.",
        "compliant": "Rewrite this compliant comment using different words while keeping it appropriate and rule-compliant.",
    },
}

2. 模型

• 通过 QLoRA 训练了三个模型：Phi-4、Qwen-2.5-14B-Instruct 和 Qwen3-14B。
• 每个模型均训练为接收 body 和 rule 作为输入，并根据 rule_violation 输出 Yes 或 No。
• 训练使用 Unsloth 配合 DDP（分布式数据并行）进行高效的多 GPU 训练。

效果不佳的尝试

• 伪标签 (pseudo label)
• muon 优化器 (muon optimizer)
• 使用 few-shot 提示生成的新训练数据 (new training data using few-shot prompting)