script/vibe/build_exp_abcd_scaled.py

import json
import os
import random
from copy import deepcopy


BASE_DIR = "/root/test/weitiao/data_process_bq/data3"
ABCD_DIR = os.path.join(BASE_DIR, "ABCD")
EXP_DIR = os.path.join(BASE_DIR, "exp")

INPUT_FILES = {
    "A": os.path.join(ABCD_DIR, "instag_A_Safety_Risk_chatml.json"),
    "B": os.path.join(ABCD_DIR, "instag_B_Romance_chatml.json"),
    "C": os.path.join(ABCD_DIR, "instag_C_Emotional_Support_chatml.json"),
    "D": os.path.join(ABCD_DIR, "instag_D_Social_Interaction_chatml.json"),
}

# 总数固定 20000；A 按 1:x 缩放；B/C/D 尽量均分后做整数分配
TARGETS = {
    "1：64": {"A": 312, "B": 6563, "C": 6563, "D": 6562},
    "1：32": {"A": 625, "B": 6459, "C": 6458, "D": 6458},
    "1：16": {"A": 1250, "B": 6250, "C": 6250, "D": 6250},
    "1：8": {"A": 2500, "B": 5834, "C": 5833, "D": 5833},
    "1：4": {"A": 5000, "B": 5000, "C": 5000, "D": 5000},
}

SEED = 20260316


def load_data() -> dict:
    data = {}
    for k, p in INPUT_FILES.items():
        if not os.path.exists(p):
            raise FileNotFoundError(f"找不到输入文件: {p}")
        with open(p, "r", encoding="utf-8") as f:
            obj = json.load(f)
        if not isinstance(obj, list):
            raise ValueError(f"输入文件顶层不是 list: {p}")
        data[k] = obj
    return data


def main():
    os.makedirs(EXP_DIR, exist_ok=True)

    data = load_data()
    print("原始各类数量：", {k: len(v) for k, v in data.items()})

    # 1) 每一类内部先打乱顺序（固定 seed，保证可复现）
    rng = random.Random(SEED)
    pools = {}
    for k, arr in data.items():
        arr_copy = deepcopy(arr)
        rng.shuffle(arr_copy)
        pools[k] = arr_copy

    # 检查是否满足所有目标
    for ratio, cnts in TARGETS.items():
        for cls, n in cnts.items():
            if n > len(pools[cls]):
                raise ValueError(f"{ratio} 目标超出 {cls} 可用数量: need={n}, have={len(pools[cls])}")
        if sum(cnts.values()) != 20000:
            raise ValueError(f"{ratio} 目标总数不是20000: {sum(cnts.values())}")

    # 2) 逐比例导出（前缀切片，保证相邻比例只做必要增减）
    for ratio, cnts in TARGETS.items():
        out_dir = os.path.join(EXP_DIR, f"exp{ratio}")
        os.makedirs(out_dir, exist_ok=True)

        merged = []
        for cls in ["A", "B", "C", "D"]:
            merged.extend(pools[cls][: cnts[cls]])

        out_path = os.path.join(out_dir, "ABCD_2w_chatml.json")
        with open(out_path, "w", encoding="utf-8") as f:
            json.dump(merged, f, ensure_ascii=False, indent=2)

        meta_path = os.path.join(out_dir, "meta.json")
        with open(meta_path, "w", encoding="utf-8") as f:
            json.dump(
                {
                    "ratio": ratio,
                    "target_counts": cnts,
                    "total": len(merged),
                    "seed": SEED,
                    "note": "each class shuffled once, then prefix-sliced for minimal-change scaling",
                },
                f,
                ensure_ascii=False,
                indent=2,
            )

        print(f"✅ {ratio} -> {out_path} | counts={cnts} | total={len(merged)}")


if __name__ == "__main__":
    main()

import json
import os
import random
from copy import deepcopy


BASE_DIR = "/root/test/weitiao/data_process_bq/data3"
ABCD_DIR = os.path.join(BASE_DIR, "ABCD")
EXP_DIR = os.path.join(BASE_DIR, "exp")

INPUT_FILES = {
    "A": os.path.join(ABCD_DIR, "instag_A_Safety_Risk_chatml.json"),
    "B": os.path.join(ABCD_DIR, "instag_B_Romance_chatml.json"),
    "C": os.path.join(ABCD_DIR, "instag_C_Emotional_Support_chatml.json"),
    "D": os.path.join(ABCD_DIR, "instag_D_Social_Interaction_chatml.json"),
}

# 总数固定 20000；A 按 1:x 缩放；B/C/D 尽量均分后做整数分配
TARGETS = {
    "1：64": {"A": 312, "B": 6563, "C": 6563, "D": 6562},
    "1：32": {"A": 625, "B": 6459, "C": 6458, "D": 6458},
    "1：16": {"A": 1250, "B": 6250, "C": 6250, "D": 6250},
    "1：8": {"A": 2500, "B": 5834, "C": 5833, "D": 5833},
    "1：4": {"A": 5000, "B": 5000, "C": 5000, "D": 5000},
}

SEED = 20260316


def load_data() -> dict:
    data = {}
    for k, p in INPUT_FILES.items():
        if not os.path.exists(p):
            raise FileNotFoundError(f"找不到输入文件: {p}")
        with open(p, "r", encoding="utf-8") as f:
            obj = json.load(f)
        if not isinstance(obj, list):
            raise ValueError(f"输入文件顶层不是 list: {p}")
        data[k] = obj
    return data


def main():
    os.makedirs(EXP_DIR, exist_ok=True)

    data = load_data()
    print("原始各类数量：", {k: len(v) for k, v in data.items()})

    # 1) 每一类内部先打乱顺序（固定 seed，保证可复现）
    rng = random.Random(SEED)
    pools = {}
    for k, arr in data.items():
        arr_copy = deepcopy(arr)
        rng.shuffle(arr_copy)
        pools[k] = arr_copy

    # 检查是否满足所有目标
    for ratio, cnts in TARGETS.items():
        for cls, n in cnts.items():
            if n > len(pools[cls]):
                raise ValueError(f"{ratio} 目标超出 {cls} 可用数量: need={n}, have={len(pools[cls])}")
        if sum(cnts.values()) != 20000:
            raise ValueError(f"{ratio} 目标总数不是20000: {sum(cnts.values())}")

    # 2) 逐比例导出（前缀切片，保证相邻比例只做必要增减）
    for ratio, cnts in TARGETS.items():
        out_dir = os.path.join(EXP_DIR, f"exp{ratio}")
        os.makedirs(out_dir, exist_ok=True)

        merged = []
        for cls in ["A", "B", "C", "D"]:
            merged.extend(pools[cls][: cnts[cls]])

        # 合并后再打乱（可注释掉；若想保持 A/B/C/D 分块顺序就不打乱）
        # random.Random(SEED).shuffle(merged)

        out_path = os.path.join(out_dir, "ABCD_2w_chatml.json")
        with open(out_path, "w", encoding="utf-8") as f:
            json.dump(merged, f, ensure_ascii=False, indent=2)

        meta_path = os.path.join(out_dir, "meta.json")
        with open(meta_path, "w", encoding="utf-8") as f:
            json.dump(
                {
                    "ratio": ratio,
                    "target_counts": cnts,
                    "total": len(merged),
                    "seed": SEED,
                    "note": "each class shuffled once, then prefix-sliced for minimal-change scaling",
                },
                f,
                ensure_ascii=False,
                indent=2,
            )

        print(f"✅ {ratio} -> {out_path} | counts={cnts} | total={len(merged)}")


if __name__ == "__main__":
    main()