script/safe_guard_count.py

import json
from collections import Counter

INPUT_FILE = "/root/test/weitiao/data_process_bq/data34/set1_extracted/B_Romance_3249_safeguard_set1.jsonl"

# conversations 中带 _safety 的 gpt 轮：Unsafe 占比 **严格大于** 该比例时，样本级 conversation 记为 Unsafe
UNSAFE_CONV_RATIO_THRESHOLD = 0.3

# ── 统计容器 ──────────────────────────────────────────────
stats = {
    "total_samples": 0,
    "conv_labels":     Counter(),
    "chosen_labels":   Counter(),
    "rejected_labels": Counter(),
    "conv_categories":     Counter(),
    "chosen_categories":   Counter(),
    "rejected_categories": Counter(),
    "sample_any_unsafe":        0,
    "sample_any_controversial": 0,
    "sample_all_safe":          0,
    "total_conv_turns": 0,
}

unsafe_turn_count        = Counter()
controversial_turn_count = Counter()
triple_counter           = Counter()

with open(INPUT_FILE, "r", encoding="utf-8") as f:
    for line in f:
        sample = json.loads(line)
        stats["total_samples"] += 1
        sample_labels = []

        # ── conversations gpt turn ────────────────────────
        n_unsafe_conv        = 0
        n_controversial_conv = 0
        conv_turn_labels     = []

        for turn in sample["conversations"]:
            if turn["from"] != "gpt" or "_safety" not in turn:
                continue
            safety = turn["_safety"]
            label  = safety.get("label")
            stats["total_conv_turns"] += 1
            stats["conv_labels"][label] += 1
            for cat in safety.get("categories", []):
                stats["conv_categories"][cat] += 1
            sample_labels.append(label)
            conv_turn_labels.append(label)
            if label == "Unsafe":
                n_unsafe_conv += 1
            elif label == "Controversial":
                n_controversial_conv += 1

        unsafe_turn_count[n_unsafe_conv] += 1
        controversial_turn_count[n_controversial_conv] += 1

        # ── chosen ────────────────────────────────────────
        chosen_label = None
        if "_safety" in sample.get("chosen", {}):
            safety       = sample["chosen"]["_safety"]
            chosen_label = safety.get("label")
            stats["chosen_labels"][chosen_label] += 1
            for cat in safety.get("categories", []):
                stats["chosen_categories"][cat] += 1
            sample_labels.append(chosen_label)

        # ── rejected ──────────────────────────────────────
        rejected_label = None
        if "_safety" in sample.get("rejected", {}):
            safety         = sample["rejected"]["_safety"]
            rejected_label = safety.get("label")
            stats["rejected_labels"][rejected_label] += 1
            for cat in safety.get("categories", []):
                stats["rejected_categories"][cat] += 1
            sample_labels.append(rejected_label)

        # ── 样本级汇总 ────────────────────────────────────
        if "Unsafe" in sample_labels:
            stats["sample_any_unsafe"] += 1
        elif "Controversial" in sample_labels:
            stats["sample_any_controversial"] += 1
        else:
            stats["sample_all_safe"] += 1

        # ── 三元组（三分类）──────────────────────────────
        # 仅统计 conversations 里 from=gpt 且含 _safety 的轮；Unsafe 轮数 / 该总轮数 > 30% 则整条为 Unsafe
        n_conv_turns = len(conv_turn_labels)
        if n_conv_turns == 0:
            conv_status = "Safe"
        elif (n_unsafe_conv / n_conv_turns) >= UNSAFE_CONV_RATIO_THRESHOLD:
            conv_status = "Unsafe"
        elif any(l == "Controversial" for l in conv_turn_labels):
            conv_status = "Controversial"
        else:
            conv_status = "Safe"

        chosen_status   = chosen_label   if chosen_label   in ("Safe", "Controversial", "Unsafe") else "Safe"
        rejected_status = rejected_label if rejected_label in ("Safe", "Controversial", "Unsafe") else "Safe"

        triple_counter[(
            f"conv:{conv_status}",
            f"chosen:{chosen_status}",
            f"rejected:{rejected_status}",
        )] += 1

# ── 打印报告 ──────────────────────────────────────────────
total = stats["total_samples"]

print(f"{'='*50}")
print(f"总样本数: {total}")
print(
    f"conversation 级 Unsafe 判定：gpt 且含 _safety 的轮中，Unsafe 轮占比 > "
    f"{UNSAFE_CONV_RATIO_THRESHOLD * 100:.0f}%"
)
print(f"{'='*50}")

# 样本级
print(f"\n── 样本级分布 ──────────────────────────────────")
print(f"  全部 Safe        : {stats['sample_all_safe']:>6}  ({stats['sample_all_safe']/total*100:.1f}%)")
print(f"  含 Controversial : {stats['sample_any_controversial']:>6}  ({stats['sample_any_controversial']/total*100:.1f}%)")
print(f"  含 Unsafe        : {stats['sample_any_unsafe']:>6}  ({stats['sample_any_unsafe']/total*100:.1f}%)")

# 各位置 label / category 分布
for name, label_counter, cat_counter, turn_total in [
    ("conversations gpt turn", stats["conv_labels"],     stats["conv_categories"],     stats["total_conv_turns"]),
    ("chosen",                 stats["chosen_labels"],   stats["chosen_categories"],   total),
    ("rejected",               stats["rejected_labels"], stats["rejected_categories"], total),
]:
    print(f"\n── {name} (共 {turn_total} 条) ─────────────────────")
    print("  Label 分布:")
    for label, cnt in label_counter.most_common():
        print(f"    {label:<15}: {cnt:>6}  ({cnt/turn_total*100:.1f}%)")
    print("  Category 分布 (不含 None):")
    for cat, cnt in cat_counter.most_common():
        if cat == "None":
            continue
        print(f"    {cat:<40}: {cnt:>6}  ({cnt/turn_total*100:.1f}%)")

# conversations unsafe/controversial turn 数量分布
print(f"\n── conversations 中 Unsafe turn 数量分布 ────────────")
print(f"  {'Unsafe turns':<15} {'样本数':>8}  {'占比':>8}  {'累计占比':>10}")
cumulative = 0
for n in sorted(unsafe_turn_count):
    cnt = unsafe_turn_count[n]
    cumulative += cnt
    print(f"  {n:<15} {cnt:>8}  {cnt/total*100:>7.1f}%  {cumulative/total*100:>9.1f}%")

print(f"\n── conversations 中 Controversial turn 数量分布 ─────")
print(f"  {'Controversial':<15} {'样本数':>8}  {'占比':>8}  {'累计占比':>10}")
cumulative = 0
for n in sorted(controversial_turn_count):
    cnt = controversial_turn_count[n]
    cumulative += cnt
    print(f"  {n:<15} {cnt:>8}  {cnt/total*100:>7.1f}%  {cumulative/total*100:>9.1f}%")

# 三元组分布（27 种组合，跳过空的）
print(f"\n── 三元组分布 (conv / chosen / rejected) ───────────")
print(f"  {'conv':<16} {'chosen':<16} {'rejected':<16} {'样本数':>8}  {'占比':>8}")
print(f"  {'-'*66}")

all_keys = [
    (c, ch, r)
    for c  in ("conv:Safe",      "conv:Controversial",      "conv:Unsafe")
    for ch in ("chosen:Safe",    "chosen:Controversial",    "chosen:Unsafe")
    for r  in ("rejected:Safe",  "rejected:Controversial",  "rejected:Unsafe")
]

for key in all_keys:
    cnt = triple_counter.get(key, 0)
    if cnt == 0:
        continue
    c, ch, r = [k.split(":")[1] for k in key]
    print(f"  {c:<16} {ch:<16} {r:<16} {cnt:>8}  ({cnt/total*100:.1f}%)")

print(f"  {'-'*66}")
print(f"  {'合计':<48} {total:>8}  (100.0%)")