script/special_phe_distribution.py

import json
from collections import Counter
import matplotlib.pyplot as plt
import numpy as np
from decimal import Decimal

path = "/root/test/weitiao/data_process_bq/data/test_filtered_by_length_scored_by_rmv2.2.json"

with open(path, "r", encoding="utf-8") as f:
    dataset = json.load(f, parse_float=Decimal)
    print(f"成功读取 {len(dataset)} 条数据")

chosen_scores = []
rejected_scores = []
delta_scores = []
records = []

for item in dataset:
    c = item.get("chosen_score")
    r = item.get("rejected_score")
    d = item.get("delta")

    if c is None or r is None or d is None:
        continue

    chosen_scores.append(c)
    rejected_scores.append(r)
    delta_scores.append(d)

    records.append({"chosen": c, "rejected": r, "delta": d})

print(f"有效样本数：{len(records)}")

min_d = int(np.floor(min(delta_scores)))
max_d = int(np.ceil(max(delta_scores)))
bins = list(range(min_d, max_d + 1))

hist_counts = Counter(int(d) for d in delta_scores)

plt.figure(figsize=(12, 6))
plt.bar(hist_counts.keys(), hist_counts.values())
plt.title("Delta Score Distribution (bin=1)")
plt.xlabel("Delta Score (integer)")
plt.ylabel("Count")
plt.grid(axis='y')
plt.savefig("/root/test/weitiao/data_process_bq/photo/delta_score_openrlhf_rmv2.2.png", dpi=200)
plt.close()

print("柱状图已保存为 delta_score_openrlhf_rmv2.2.png") 

def top3(counter):
    return counter.most_common(3)

chosen_counter = Counter(chosen_scores)
rejected_counter = Counter(rejected_scores)
delta_counter = Counter(delta_scores)

print("\n=== Top 3 values for each score field ===")
print("chosen_score:", top3(chosen_counter))
print("rejected_score:", top3(rejected_counter))
print("delta_score:", top3(delta_counter))

print("\n=== Overlap Analysis ===")

# delta_score 出现最多的值
delta_top_value, delta_top_count = delta_counter.most_common(1)[0]

# 筛选出这些记录
subset = [rec for rec in records if rec["delta"] == delta_top_value]

subset_chosen_counter = Counter(rec["chosen"] for rec in subset)
subset_rejected_counter = Counter(rec["rejected"] for rec in subset)

chosen_in_subset, chosen_cnt = subset_chosen_counter.most_common(1)[0]
rejected_in_subset, rejected_cnt = subset_rejected_counter.most_common(1)[0]

total_subset = len(subset)

print(f"delta_score 最常出现的值: {delta_top_value} (共 {delta_top_count} 条)")
print(f"  -> 在这些样本中 chosen_score 最常出现: {chosen_in_subset} "
      f"({chosen_cnt} 次，占比 {chosen_cnt/total_subset:.2%})")
print(f"  -> 在这些样本中 rejected_score 最常出现: {rejected_in_subset} "
      f"({rejected_cnt} 次，占比 {rejected_cnt/total_subset:.2%})")