script/safe_guard.py

import os
import json
import torch
from tqdm import tqdm
from transformers import AutoTokenizer, AutoModelForCausalLM
import re
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

# =========================
# 配置
# =========================
MODEL_ID = "meta-llama/Llama-Guard-3-8B"
# --- 核心修改：输入现在是单个文件名 ---
INPUT_FILENAME = "/root/test/weitiao/data_processing_hsichen/data_process_bq/data/rm_dpo_infered_8_truncated_replaced.json" 
# 👇 输出依然是一个目录，用于存放 8 个象限文件
OUTPUT_DIRECTORY = "/root/test/weitiao/data_processing_hsichen/data_process_bq/result/high_chosen_reward_safety_8_quadrants_labeled"
PLOT_FILENAME = "high_chosen_reward_safety_8_quadrants_distribution_labeled.png"

device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float16

HISTORY_UNSAFE_RATIO_THRESHOLD = 0.20 
HAZARD_CATEGORY_MAP = {
    "S1": "Violent Crimes", "S2": "Non-Violent Crimes", "S3": "Sex-Related Crimes",
    "S4": "Child Sexual Exploitation", "S5": "Defamation", "S6": "Specialized Advice",
    "S7": "Privacy", "S8": "Intellectual Property", "S9": "Indiscriminate Weapons",
    "S10": "Hate", "S11": "Suicide & Self-Harm", "S12": "Sexual Content",
    "S13": "Elections", "S14": "Code Interpreter Abuse"
}
ALL_UNSAFE_LABELS = set(HAZARD_CATEGORY_MAP.keys())

# --- 加载模型 ---
print(f"🚀 Loading Safety Model: {MODEL_ID}")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, trust_remote_code=True, torch_dtype=torch_dtype, device_map=device)
model.eval()

# =========================
# 辅助函数 (保持不变)
# =========================
def moderate_and_get_unsafe_labels(text):
    if not text or not text.strip(): return False, set()
    chat = [{"role": "user", "content": text}]
    inputs = tokenizer.apply_chat_template(chat, return_tensors="pt").to(device)
    output = model.generate(input_ids=inputs, max_new_tokens=100, pad_token_id=tokenizer.eos_token_id, do_sample=False)
    result_text = tokenizer.decode(output[0, -100:], skip_special_tokens=True)
    if "assistant\n" in result_text:
        result_text = result_text.split("assistant\n")[-1].strip()
    if "unsafe" in result_text:
        labels = set(re.findall(r'S\d+', result_text))
        violated_labels = labels.intersection(ALL_UNSAFE_LABELS)
        if violated_labels:
            return True, violated_labels
    return False, set()

def plot_quadrant_distribution(quadrant_counts, plot_filename):
    print(f"\n📊 正在绘制安全象限分布图...")
    labels = list(quadrant_counts.keys())
    counts = list(quadrant_counts.values())
    sns.set_theme(style="whitegrid")
    plt.figure(figsize=(12, 8))
    ax = sns.barplot(x=counts, y=labels, palette="coolwarm", orient='h')
    ax.set_title('Distribution of DPO Data by Safety Quadrants', fontsize=16, pad=20)
    ax.set_xlabel('Number of Samples', fontsize=12)
    ax.set_ylabel('Safety Quadrant (History / Chosen / Rejected)', fontsize=12)
    for index, value in enumerate(counts):
        ax.text(value, index, f' {value}', va='center', fontsize=10)
    plt.tight_layout()
    plt.savefig(plot_filename)
    print(f"✅ 分布图已保存到 '{plot_filename}'。")
    plt.close()

# =========================
# 主流程 (已简化为处理单个文件)
# =========================
def classify_and_label_single_file(input_file, output_dir, history_threshold):
    # --- 1. 加载数据 ---
    print(f"正在加载 DPO 数据: {input_file}...")

    with open(input_file, 'r', encoding='utf-8') as f:
        data = json.load(f)


    # --- 2. 初始化 ---
    if not os.path.exists(output_dir):
        print(f"正在创建输出目录: {output_dir}")
        os.makedirs(output_dir)

    quadrants = {name: [] for name in [
        "history_safe__chosen_safe__rejected_safe", "history_safe__chosen_safe__rejected_unsafe",
        "history_safe__chosen_unsafe__rejected_safe", "history_safe__chosen_unsafe__rejected_unsafe",
        "history_unsafe__chosen_safe__rejected_safe", "history_unsafe__chosen_safe__rejected_unsafe",
        "history_unsafe__chosen_unsafe__rejected_safe", "history_unsafe__chosen_unsafe__rejected_unsafe"]}

    # --- 3. 遍历并分类数据 ---
    for item in tqdm(data, desc=f"分类 {os.path.basename(input_file)}"):
        # (分类和标注的核心逻辑与之前完全相同)
        history_unsafe_count, all_history_labels = 0, set()
        gpt_turns = [t.get("value", "").strip() for t in item.get("conversations", []) if t.get("from") == "gpt"]
        if gpt_turns:
            for text in gpt_turns:
                is_unsafe, labels = moderate_and_get_unsafe_labels(text)
                if is_unsafe: history_unsafe_count += 1; all_history_labels.update(labels)
        is_history_unsafe = (history_unsafe_count / len(gpt_turns)) > history_threshold if gpt_turns else False

        is_chosen_unsafe, chosen_labels = moderate_and_get_unsafe_labels(item.get("chosen", {}).get("value", ""))
        is_rejected_unsafe, rejected_labels = moderate_and_get_unsafe_labels(item.get("rejected", {}).get("value", ""))

        history_status = "unsafe" if is_history_unsafe else "safe"
        chosen_status = "unsafe" if is_chosen_unsafe else "safe"
        rejected_status = "unsafe" if is_rejected_unsafe else "safe"
        quadrant_name = f"history_{history_status}__chosen_{chosen_status}__rejected_{rejected_status}"

        if history_status == 'unsafe' or chosen_status == 'unsafe' or rejected_status == 'unsafe':
            item['safety_analysis'] = {
                "history_status": history_status,
                "history_labels": [HAZARD_CATEGORY_MAP.get(c, c) for c in sorted(list(all_history_labels))],
                "chosen_status": chosen_status,
                "chosen_labels": [HAZARD_CATEGORY_MAP.get(c, c) for c in sorted(list(chosen_labels))],
                "rejected_status": rejected_status,
                "rejected_labels": [HAZARD_CATEGORY_MAP.get(c, c) for c in sorted(list(rejected_labels))]
            }
        quadrants[quadrant_name].append(item)

    # --- 4. 保存文件并统计 ---
    print("\n" + "="*30 + " 处理完成 - 正在保存结果 " + "="*30)
    quadrant_counts = {name: len(data) for name, data in quadrants.items()}
    for quadrant_name, quadrant_data in quadrants.items():
        output_path = os.path.join(output_dir, f"{quadrant_name}.json")
        with open(output_path, 'w', encoding='utf-8') as f: json.dump(quadrant_data, f, indent=2, ensure_ascii=False)
        print(f"  - {quadrant_name}: {len(quadrant_data):>5} 条 -> 已保存")

    # --- 5. 绘制分布图 ---
    plot_quadrant_distribution(quadrant_counts, PLOT_FILENAME)
    total_processed = sum(quadrant_counts.values())
    print(f"\n 任务完成。总共处理并分类了 {total_processed} 条数据。")
    print(f" 所有文件都已保存在 '{output_dir}' 目录中。")


if __name__ == "__main__":
    classify_and_label_single_file(INPUT_FILENAME, OUTPUT_DIRECTORY, HISTORY_UNSAFE_RATIO_THRESHOLD)