script/ngram_for_repeated.py

import json
import re
from collections import Counter
from itertools import islice

# === 配置 ===
filename = "/root/test/weitiao/data_processing_hsichen/data_process_bq/data/mistral_group5_redpo_pk.json"        #pk语料下载下来的直接转json格式
model_id = 7746        #模型id
n = 5        # n-gram 短语长度（3表示统计三连词短语）
top_n = 30   # 输出最常见的前多少个短语

# === 清理与分词函数 ===
def clean_and_tokenize(text):
    text = text.lower()
    text = re.sub(r'\d+', '', text)  # 去掉数字
    text = re.sub(r'[^a-z\s]', '', text)  # 去掉非字母字符
    text = re.sub(r'\s+', ' ', text).strip()
    return text.split()

# === 生成 n-gram ===
def get_ngrams(tokens, n):
    if len(tokens) < n:
        return []
    return [' '.join(tokens[i:i+n]) for i in range(len(tokens)-n+1)]

# === 计数容器 ===
chosen_phrases = []
reject_phrases = []

# === 读取 JSON 文件 ===
with open(filename, 'r', encoding='utf-8') as f:
    data = json.load(f)

# === 遍历数据 ===
for item in data:
    chosen_model = item.get("chosen_model")
    reject_model = item.get("reject_model")
    chosen_text = item.get("chosen", "")
    reject_text = item.get("reject", "")

    # 统计 chosen
    if chosen_model == model_id and isinstance(chosen_text, str):
        tokens = clean_and_tokenize(chosen_text)
        chosen_phrases.extend(get_ngrams(tokens, n))

    # 统计 reject
    if reject_model == model_id and isinstance(reject_text, str):
        tokens = clean_and_tokenize(reject_text)
        reject_phrases.extend(get_ngrams(tokens, n))

# === 统计 n-gram 频率 ===
chosen_counter = Counter(chosen_phrases)
reject_counter = Counter(reject_phrases)

# === 打印结果 ===
print(f"\n=== 模型ID: {model_id} | {n}-gram 高频短语统计 ===")
print(f"共分析 {len(chosen_phrases)} 个 chosen 短语，{len(reject_phrases)} 个 reject 短语。")

print(f"\n chosen_model == {model_id} 的前 {top_n} 个高频短语:")
for phrase, count in islice(chosen_counter.most_common(top_n), top_n):
    print(f"{count:>5} ×  {phrase}")

print(f"\n reject_model == {model_id} 的前 {top_n} 个高频短语:")
for phrase, count in islice(reject_counter.most_common(top_n), top_n):
    print(f"{count:>5} ×  {phrase}")