#!/usr/bin/env python
# -*- coding: utf-8 -*-
# 汇总所有模型在 HumanEval+/MBPP+/LiveCodeBench v6 上的评测结果, 输出到同一个 TXT
# 自动扫描:
#   evalplus_results/{humaneval,mbpp}/*_eval_results.json  (文件名含模型路径标识)
#   lcb_results/<模型名>/Scenario.codegeneration_*_eval_all.json
# 用法:
#   python summarize_results.py            # 汇总全部模型 -> summary_all.txt
#   python summarize_results.py 模型A 模型B  # 只汇总指定模型
# 模型名 = LCB 结果目录名 (也是 EvalPlus 标识的最后一段)

import json
import sys
from datetime import datetime
from pathlib import Path

BASE = Path("/data/icto/mc56499/code_eval")
EVALPLUS_DIR = BASE / "evalplus_results"
LCB_DIR = BASE / "lcb_results"
OUT = BASE / "summary_all.txt"


def evalplus_stats(path: Path):
    """返回 (base_pass, plus_pass, n_tasks)"""
    with open(path) as f:
        r = json.load(f)
    evals = r["eval"]
    n = len(evals)
    if n == 0:
        return None
    base = sum(1 for ss in evals.values() if any(s["base_status"] == "pass" for s in ss))
    plus = sum(
        1
        for ss in evals.values()
        if any(s["base_status"] == "pass" and s["plus_status"] == "pass" for s in ss)
    )
    return base / n, plus / n, n


def lcb_stats(path: Path):
    """返回 dict: overall/easy/medium/hard pass@1 及题数"""
    with open(path) as f:
        results = json.load(f)
    if not results:
        return None
    by_diff = {}
    for r in results:
        by_diff.setdefault(r["difficulty"], []).append(r["pass@1"])
    stats = {"n": len(results), "overall": sum(r["pass@1"] for r in results) / len(results)}
    for d in ["easy", "medium", "hard"]:
        if d in by_diff:
            vals = by_diff[d]
            stats[d] = sum(vals) / len(vals)
            stats[f"{d}_n"] = len(vals)
    return stats


def discover_models():
    """扫描两个结果目录, 以"模型名"为 key 合并 (EvalPlus 标识最后一段 = LCB 目录名)"""
    models = {}
    # EvalPlus: 文件名形如 <路径标识>_vllm_temp_0.0_eval_results.json, 路径标识中 / 换成 --
    for dataset in ["humaneval", "mbpp"]:
        d = EVALPLUS_DIR / dataset
        if not d.is_dir():
            continue
        for f in sorted(d.glob("*_eval_results.json")):
            ident = f.name.split("_vllm_temp_")[0]  # 去掉 _vllm_temp_0.0_eval_results.json
            name = ident.split("--")[-1]  # 取最后一段作为模型名
            m = models.setdefault(name, {})
            st = evalplus_stats(f)
            if st:
                m[f"{dataset}_base"], m[f"{dataset}_plus"], m[f"{dataset}_n"] = st
    # LCB
    if LCB_DIR.is_dir():
        for d in sorted(LCB_DIR.iterdir()):
            if not d.is_dir():
                continue
            files = sorted(d.glob("Scenario.codegeneration_*_eval_all.json"))
            if not files:
                continue
            m = models.setdefault(d.name, {})
            st = lcb_stats(files[-1])  # 最新一次
            if st:
                m.update({f"lcb_{k}": v for k, v in st.items()})
    return models


def fmt(v):
    return f"{v * 100:.2f}" if v is not None else "--"


def main():
    selected = sys.argv[1:]
    models = discover_models()
    if selected:
        models = {k: v for k, v in models.items() if k in selected}
    if not models:
        print("未找到任何评测结果")
        return

    L = []
    L.append("=" * 96)
    L.append("  代码评测汇总报告 (全部实验)")
    L.append("  基准: HumanEval+/MBPP+ (EvalPlus, greedy, pass@1) + LiveCodeBench v6 (n=1, temp=0.2, pass@1)")
    L.append(f"  生成时间: {datetime.now().strftime('%Y-%m-%d %H:%M')}")
    L.append("=" * 96)

    # ---------- 一、总览对比 ----------
    L.append("")
    L.append("一、总览对比表 (数字为 pass@1 %, -- 表示无结果)")
    L.append("-" * 96)
    w = max(34, max(len(k) for k in models) + 2)
    header = (
        f"  {'模型':<{w}}"
        f"{'HE+base':>9}{'HE+plus':>9}{'MBPPbase':>10}{'MBPPplus':>10}"
        f"{'LCB-all':>9}{'LCB-easy':>10}{'LCB-med':>9}{'LCB-hard':>10}"
    )
    L.append(header)
    L.append("  " + "-" * (w + 76))
    for name, m in sorted(models.items(), key=lambda x: -(x[1].get("lcb_overall") or -1)):
        row = (
            f"  {name:<{w}}"
            f"{fmt(m.get('humaneval_base')):>9}{fmt(m.get('humaneval_plus')):>9}"
            f"{fmt(m.get('mbpp_base')):>10}{fmt(m.get('mbpp_plus')):>10}"
            f"{fmt(m.get('lcb_overall')):>9}{fmt(m.get('lcb_easy')):>10}"
            f"{fmt(m.get('lcb_medium')):>9}{fmt(m.get('lcb_hard')):>10}"
        )
        L.append(row)
    L.append("")

    # ---------- 二、分模型明细 ----------
    L.append("二、分模型明细")
    for name, m in sorted(models.items()):
        L.append("-" * 96)
        L.append(f"◆ {name}")
        # EvalPlus
        if "humaneval_n" in m or "mbpp_n" in m:
            L.append(f"  [EvalPlus greedy pass@1]")
            for ds, dname in [("humaneval", "HumanEval+"), ("mbpp", "MBPP+")]:
                if f"{ds}_n" in m:
                    L.append(
                        f"    {dname:<12} base {fmt(m[f'{ds}_base'])}%"
                        f"   plus {fmt(m[f'{ds}_plus'])}%"
                        f"   (通过 {round(m[f'{ds}_plus'] * m[f'{ds}_n'])}/{m[f'{ds}_n']})"
                    )
        else:
            L.append("  [EvalPlus] 无结果")
        # LCB
        if "lcb_n" in m:
            L.append(f"  [LiveCodeBench v6] n=1 temp=0.2 pass@1, 共 {m['lcb_n']} 题")
            L.append(f"    Overall  {fmt(m['lcb_overall'])}%")
            for d in ["easy", "medium", "hard"]:
                if f"lcb_{d}" in m:
                    L.append(f"    {d:<8} {fmt(m[f'lcb_{d}'])}%  ({m[f'lcb_{d}_n']}题)")
        else:
            L.append("  [LiveCodeBench] 无结果")
    L.append("=" * 96)

    text = "\n".join(L)
    OUT.write_text(text, encoding="utf-8")
    print(text)
    print(f"\n[已保存] {OUT}  (共 {len(models)} 个模型)")


if __name__ == "__main__":
    main()