evaluation/vstyle/scoring/call_api.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import argparse
import base64
import json
import os
import re
import time
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
from typing import Any, Dict, List, Tuple

from openai import OpenAI
from tqdm import tqdm

try:
    import matplotlib.pyplot as plt
except ImportError:
    plt = None

try:
    import pandas as pd
except ImportError:
    pd = None

if plt is not None:
    plt.rcParams.update({"font.size": 10})

SUB_CATS = {
    "acoustic_attributes": [
        "acoustic_attributes/age",
        "acoustic_attributes/speed",
        "acoustic_attributes/gender",
        "acoustic_attributes/emotion",
        "acoustic_attributes/pitch",
        "acoustic_attributes/volume",
        "acoustic_attributes/composite_properties",
    ],
    "instruction": [
        "instruction/emotion",
        "instruction/variation",
        "instruction/style",
    ],
    "role_play": [
        "role_play/character",
        "role_play/scenario",
    ],
    "empathy": [
        "empathy/anger",
        "empathy/sadness_disappointment",
        "empathy/anxiety_fear",
        "empathy/joy_excitement",
    ],
}


# --------------------------- Utility Functions --------------------------------

def load_prompts(prompts_dir: Path) -> Dict[str, str]:
    mapping = {}
    for file in prompts_dir.glob("*.txt"):
        mapping[file.stem] = file.read_text(encoding="utf-8")
    if not mapping:
        raise RuntimeError(f"No *.txt prompt files found in {prompts_dir}")
    return mapping


def construct_prompt(template: str, instruction: str, ability: str) -> str:
    return template.format(instruction_type=ability, input_instruction=instruction)


def debug_print_input(sample, root_dir, prompts):
    """跑之前调用这个，肉眼确认输入"""
    big_cat = sample["ability"].split("/")[0]
    prompt_text = construct_prompt(
        prompts[big_cat],
        instruction=sample.get("instruct_text", ""),
        ability=sample["ability"],
    )
    audio_path = Path(sample.get("final_audio_path") or root_dir / sample.get("response_audio_path", ""))

    print("=" * 60)
    print(f"[id]           {sample['id']}")
    print(f"[ability]      {sample['ability']}")
    print(f"[instruct_text]{sample['instruct_text']}")
    print(f"[audio_path]   {audio_path}  exists={audio_path.exists()}")
    print(f"[prompt末尾50字]{prompt_text[-200:]}")
    print("=" * 60)


def ensure_dir(p: Path):
    p.mkdir(parents=True, exist_ok=True)


def safe_float(x: Any):
    try:
        return float(x)
    except (TypeError, ValueError):
        return None


def normalize_score_str(x: Any) -> str:
    v = safe_float(x)
    if v is None:
        return ""
    if 1 <= v <= 5:
        if float(v).is_integer():
            return str(int(v))
        return str(v).rstrip("0").rstrip(".")
    return ""


def parse_score(model_reply: str) -> str:
    if not model_reply:
        return ""

    text = str(model_reply).strip()

    # 1) 优先匹配 [[4]] / [[4.0]]
    matches = re.findall(r"\[\[\s*([0-9]+(?:\.[0-9]+)?)\s*]]", text, flags=re.S)
    for s in reversed(matches):
        score = normalize_score_str(s)
        if score:
            return score

    # 2) 整体就是数字
    score = normalize_score_str(text)
    if score:
        return score

    # 3) 只匹配明确的最终分数字段，避免把 0-100 的风格强度误读成 1-5 分
    matches = re.findall(
        r"(?:final\s+score|score|rating|gemini_score|分数|评分|最终分数)"
        r"\s*[:：=]?\s*\[?\[?\s*([1-5](?:\.[0-9]+)?)(?!\d)\s*\]?\]?",
        text,
        flags=re.I,
    )
    for s in reversed(matches):
        score = normalize_score_str(s)
        if score:
            return score

    return ""


def encode_audio_base64(audio_path: Path) -> str:
    with audio_path.open("rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")


def build_retry_prompt(original_prompt: str, attempt: int) -> str:
    return f"""
{original_prompt}

IMPORTANT FORMAT REQUIREMENT:
Your previous response did not contain a valid score.

You MUST now output exactly ONE score between 1 and 5.
Do not explain.
Do not output any extra words.
Preferred format: [[score]]

Valid examples:
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]

This is retry attempt #{attempt}.
""".strip()


def normalize_id(value: Any):
    if value is None:
        return None
    try:
        return int(value)
    except (TypeError, ValueError):
        return str(value)


def load_dataset_metadata(dataset_metadata_path: Path) -> Dict[Any, Dict[str, Any]]:
    dataset_map: Dict[Any, Dict[str, Any]] = {}
    if not dataset_metadata_path or not Path(dataset_metadata_path).exists():
        return dataset_map

    with Path(dataset_metadata_path).open(encoding="utf-8") as f:
        for line in f:
            try:
                d = json.loads(line)
            except json.JSONDecodeError:
                continue

            instruct_id = d.get("instruct_id", d.get("id"))
            item = {
                "id": d.get("id", instruct_id),
                "instruct_id": instruct_id,
                "ability": d.get("ability", ""),
                "instruct_text": d.get("instruct_text") or d.get("audio_content", ""),
                "response_audio_path": d.get("response_audio_path") or d.get("file_name", ""),
            }
            for key in (d.get("id"), instruct_id):
                norm = normalize_id(key)
                if norm is not None:
                    dataset_map[norm] = item

    return dataset_map


def candidate_path(root_dir: Path, path_str: Any) -> Path:
    if not path_str:
        return None
    path = Path(str(path_str).strip())
    if path.is_absolute():
        return path
    return root_dir / path


def resolve_audio_path(root_dir: Path, sample: Dict[str, Any], dataset_entry: Dict[str, Any] = None) -> str:
    iid = sample.get("instruct_id", sample.get("id"))
    candidates = []

    for key in ("final_audio_path", "response_audio_path", "file_name", "response_wav"):
        path = candidate_path(root_dir, sample.get(key))
        if path:
            candidates.append(path)

    if dataset_entry:
        path = candidate_path(root_dir, dataset_entry.get("response_audio_path"))
        if path:
            candidates.append(path)

    if iid is not None:
        candidates.append(root_dir / f"{iid}.wav")

    seen = set()
    for path in candidates:
        key = str(path)
        if key in seen:
            continue
        seen.add(key)
        if path.exists():
            return str(path)
    return ""


# --------------------------- Proxy Gemini (OpenAI-compatible) -----------------

def call_proxy_gemini_api(
    prompt_text: str,
    response_audio_path: Path,
    model_name: str,
    max_retry: int,
    sleep_between_retry: int,
    api_key: str,
    base_url: str,
    temperature: float = 1.0,
    top_p: float = 0.7,
    max_tokens: int = 4096,
    verbose: bool = False,
) -> Tuple[bool, str, str]:

    client = OpenAI(api_key=api_key, base_url=base_url)
    last_err_msg = ""
    last_reply_text = ""
    audio_b64 = encode_audio_base64(response_audio_path)
    audio_format = response_audio_path.suffix.replace(".", "").lower() or "wav"

    for attempt in range(1, max_retry + 1):
        try:
            current_prompt = prompt_text if attempt == 1 else build_retry_prompt(prompt_text, attempt)

            content = [{"type": "text", "text": current_prompt}]
            content.append(
                {
                    "type": "input_audio",
                    "input_audio": {
                        "data": audio_b64,
                        "format": audio_format,
                    },
                }
            )

            resp = client.chat.completions.create(
                model=model_name,
                messages=[{"role": "user", "content": content}],
                max_tokens=max_tokens,
                temperature=temperature,
                top_p=top_p,
            )

            if not hasattr(resp, "choices") or not resp.choices:
                last_err_msg = f"Proxy API bad response: {resp}"
                print(f"[Attempt {attempt}/{max_retry}] bad response: {last_err_msg}")
            else:
                reply_text = resp.choices[0].message.content or ""
                last_reply_text = reply_text
                score_str = parse_score(reply_text)

                if verbose:
                    print(f"[Attempt {attempt}/{max_retry}] raw reply: {reply_text[:200]}")

                if score_str:
                    return True, reply_text, score_str

                last_err_msg = f"Unable to parse score from reply: {reply_text[:200]}"
                print(f"[Attempt {attempt}/{max_retry}] parse failed")

        except Exception as e:
            last_err_msg = str(e)
            print(f"[Attempt {attempt}/{max_retry}] API call error: {last_err_msg}")

        if attempt < max_retry:
            time.sleep(sleep_between_retry)

    return False, (last_reply_text or last_err_msg), ""


# --------------------------- Evaluation Core --------------------------------

def evaluate_one(
    sample: dict,
    root_dir: Path,
    prompts: Dict[str, str],
    out_dir: Path,
    model_name: str,
    max_retry: int,
    sleep_between_retry: int,
    api_key: str,
    base_url: str,
    fallback_score: str,
    temperature: float,
    top_p: float,
    max_tokens: int,
    verbose: bool,
) -> dict:
    ability = sample["ability"]
    big_cat, small_cat = ability.split("/", 1)
    sample_model_name = sample.get("model_name", "model_eval")
    sample_id = sample.get("id") or sample.get("instruct_id") or "unknown_id"
    sample["id"] = sample_id

    if big_cat not in prompts:
        error_msg = f"Prompt template not found for category {big_cat}"
        sample["gemini_score"] = fallback_score
        sample["gemini_status"] = "fallback_prompt_missing"
        sample["gemini_error"] = error_msg
        sample["gemini_raw"] = error_msg
        return sample

    response_audio_path_str = sample.get("final_audio_path", "")

    if not response_audio_path_str:
        error_msg = f"Audio file not found for ID {sample_id} in any expected location."
        sample["gemini_score"] = fallback_score
        sample["gemini_status"] = "fallback_audio_not_found"
        sample["gemini_error"] = error_msg
        sample["gemini_raw"] = error_msg
        return sample

    response_audio_path = Path(response_audio_path_str)  # 此时已经是探测好的绝对路径了

    if not response_audio_path.exists():
        error_msg = f"Audio file not found: {response_audio_path}"
        sample["gemini_score"] = fallback_score
        sample["gemini_status"] = "fallback_audio_not_found"
        sample["gemini_error"] = error_msg
        sample["gemini_raw"] = error_msg
        return sample

    prompt_text = construct_prompt(
        prompts[big_cat],
        instruction=sample.get("audio_content", "") or sample.get("instruct_text", ""),
        ability=sample["ability"],
    )

    success, reply, score = call_proxy_gemini_api(
        prompt_text=prompt_text,
        response_audio_path=response_audio_path,
        model_name=model_name,
        max_retry=min(max_retry, 5),
        sleep_between_retry=sleep_between_retry,
        api_key=api_key,
        base_url=base_url,
        temperature=temperature,
        top_p=top_p,
        max_tokens=max_tokens,
        verbose=verbose,
    )
    if success:
        sample["gemini_score"] = score
        sample["gemini_raw"] = reply
        sample["gemini_status"] = "success"
    else:
        sample["gemini_score"] = fallback_score
        sample["gemini_raw"] = reply
        sample["gemini_status"] = "fallback_after_5_failures"
        sample["gemini_error"] = f"Failed after {min(max_retry, 5)} attempts"

    # ← debug 日志必须放在这里，不能放在上面
    debug_path = out_dir / "debug_inputs_outputs.txt"
    with debug_path.open("a", encoding="utf-8") as dbg:
        dbg.write("=" * 80 + "\n")
        dbg.write(f"[ID]         {sample_id}\n")
        dbg.write(f"[ABILITY]    {ability}\n")
        dbg.write(f"[AUDIO]      {response_audio_path}\n")
        dbg.write(f"[INSTRUCT]   {sample.get('instruct_text', '')}\n")
        dbg.write(f"[PROMPT]\n{prompt_text}\n")
        dbg.write(f"[RAW OUTPUT]\n{sample.get('gemini_raw', '')}\n")
        dbg.write(f"[SCORE]      {sample.get('gemini_score', '')}\n")
        dbg.write(f"[STATUS]     {sample.get('gemini_status', '')}\n")
        dbg.write("=" * 80 + "\n\n")

    save_path = (
        out_dir
        / sample_model_name
        / "gemini_proxy_res"
        / big_cat
        / small_cat
        / f"{sample_id}.txt"
    )
    ensure_dir(save_path.parent)
    save_path.write_text(sample.get("gemini_raw", ""), encoding="utf-8")
    return sample


def load_processed_ids_and_clean_failures(scored_jsonl: Path, overwrite: bool) -> Tuple[set, int]:
    processed = set()
    cleaned_count = 0

    if not scored_jsonl.exists() or overwrite:
        return processed, cleaned_count

    valid_lines = []
    total_lines = 0

    with scored_jsonl.open("r", encoding="utf-8") as f:
        for line in f:
            total_lines += 1
            line = line.strip()
            if not line:
                continue

            try:
                rec = json.loads(line)
                gemini_score = rec.get("gemini_score", "")
                if safe_float(gemini_score) is not None:
                    processed.add(normalize_id(rec.get("id")))
                    valid_lines.append(line)
                else:
                    cleaned_count += 1
                    print(f"🗑️ Removing failed record ID {rec.get('id')}: {gemini_score}")
            except Exception as e:
                cleaned_count += 1
                print(f"🗑️ Removing invalid JSON line: {str(e)}")

    if cleaned_count > 0:
        print(f"📝 Cleaning {cleaned_count} failed records from {scored_jsonl}")
        with scored_jsonl.open("w", encoding="utf-8") as f:
            for line in valid_lines:
                f.write(line + "\n")
        print(f"✅ Cleaned file saved. Kept {len(valid_lines)} valid records out of {total_lines} total lines.")

    return processed, cleaned_count


def remove_duplicates_from_tasks(tasks: List[dict]) -> Tuple[List[dict], Dict[str, int]]:
    seen_ids = set()
    unique_tasks = []
    duplicate_count = 0
    duplicate_details = defaultdict(int)

    for task in tasks:
        task_id = normalize_id(task.get("id"))
        ability = task.get("ability", "unknown")

        if task_id not in seen_ids:
            seen_ids.add(task_id)
            unique_tasks.append(task)
        else:
            duplicate_count += 1
            duplicate_details[ability] += 1
            print(f"🔄 Removing duplicate sample ID: {task_id} (ability: {ability})")

    stats = {
        "total_duplicates": duplicate_count,
        "by_ability": dict(duplicate_details),
    }
    return unique_tasks, stats


def eval_all_samples(
    root_dir: Path,
    metadata_path: Path,
    prompts_dir: Path,
    out_dir: Path,
    max_per_ability: int,
    concurrency: int,
    overwrite: bool,
    model_name: str,
    max_retry: int,
    sleep_between_retry: int,
    api_key: str,
    base_url: str,
    fallback_score: str,
    dataset_metadata_path: Path = None,
    temperature: float = 1.0,
    top_p: float = 0.7,
    max_tokens: int = 4096,
    verbose: bool = False,
) -> Path:
    prompts = load_prompts(prompts_dir)

    dataset_map = load_dataset_metadata(dataset_metadata_path) if dataset_metadata_path else {}
    if dataset_map:
        print(f"📚 Loaded {len(dataset_map)} entries from dataset: {dataset_metadata_path}")

    scored_jsonl = out_dir / "metadata_with_score.jsonl"
    processed_ids, _ = load_processed_ids_and_clean_failures(scored_jsonl, overwrite)

    tasks: List[dict] = []
    per_ability_counter = defaultdict(int)
    skipped_missing_dataset = 0
    skipped_missing_required = 0

    print(f"📖 Reading samples from: {metadata_path}")
    with metadata_path.open(encoding="utf-8") as fin:
        for line in fin:
            try:
                sample = json.loads(line)
            except json.JSONDecodeError:
                continue

            iid = sample.get("instruct_id", sample.get("id"))
            iid_norm = normalize_id(iid)
            dataset_entry = dataset_map.get(iid_norm) if dataset_map else None

            # 传入官方 dataset metadata 时，用官方文本/类别覆盖 result.jsonl，
            # 这样评分严格对齐 VStyle 的 ground-truth instruction。
            if dataset_map:
                if not dataset_entry:
                    skipped_missing_dataset += 1
                    continue
                entry = dataset_entry
                sample["id"] = entry["id"]
                sample["instruct_id"] = entry["instruct_id"]
                sample["ability"] = entry["ability"]
                sample["instruct_text"] = entry["instruct_text"]
            else:
                sample["id"] = sample.get("id", iid)
                sample["instruct_id"] = sample.get("instruct_id", iid)
                sample["instruct_text"] = sample.get("instruct_text") or sample.get("audio_content", "")

            if not sample.get("id") or not sample.get("ability") or not sample.get("instruct_text"):
                skipped_missing_required += 1
                continue

            sample["final_audio_path"] = resolve_audio_path(root_dir, sample, dataset_entry)
            if not str(sample.get("model_name", "")).strip():
                sample["model_name"] = "unknown_model"
            ability = sample["ability"]

            if normalize_id(sample["id"]) in processed_ids:
                continue
            if per_ability_counter[ability] >= max_per_ability:
                continue

            per_ability_counter[ability] += 1
            tasks.append(sample)

    print(f"📊 Step 1 - After initial filtering: {len(tasks)} samples")
    if skipped_missing_dataset:
        print(f"📊 Skipped {skipped_missing_dataset} samples not found in dataset metadata")
    if skipped_missing_required:
        print(f"📊 Skipped {skipped_missing_required} samples missing id/ability/instruction")
    print(f"📊 Step 2 - Before deduplication: {len(tasks)} samples")

    tasks, dup_stats = remove_duplicates_from_tasks(tasks)
    print(f"📊 Step 3 - After deduplication: {len(tasks)} samples (removed {dup_stats['total_duplicates']} duplicates)")

    # for s in tasks[:3]:
    #     debug_print_input(s, root_dir, prompts)
    # input("按回车继续全量评测...")

    if not tasks:
        print("\n✅ No new samples to process, program ended.")
        return scored_jsonl

    ensure_dir(scored_jsonl.parent)
    write_mode = "a" if scored_jsonl.exists() and not overwrite else "w"

    print(f"\n🎯 Starting evaluation of {len(tasks)} samples...")
    print("=" * 80)

    with (
        scored_jsonl.open(write_mode, encoding="utf-8") as fout,
        ThreadPoolExecutor(max_workers=concurrency) as executor,
        tqdm(total=len(tasks), desc="Evaluating") as pbar,
    ):
        futures = {
            executor.submit(
                evaluate_one,
                s,
                root_dir,
                prompts,
                out_dir,
                model_name,
                max_retry,
                sleep_between_retry,
                api_key,
                base_url,
                fallback_score,
                temperature,
                top_p,
                max_tokens,
                verbose,
            ): s
            for s in tasks
        }

        for fut in as_completed(futures):
            res = fut.result()
            fout.write(json.dumps(res, ensure_ascii=False) + "\n")
            fout.flush()
            pbar.update(1)

    print(f"\n✅ All processing complete, wrote {len(tasks)} new results -> {scored_jsonl}")
    return scored_jsonl


# --------------------------- Analysis and Visualization ------------------------------

def get_ordered_abilities():
    ordered_abilities = []
    for big_cat in ["acoustic_attributes", "instruction", "role_play", "empathy"]:
        ordered_abilities.extend(SUB_CATS[big_cat])
    return ordered_abilities


def analyze_scores(scored_jsonl: Path, out_dir: Path):
    if pd is None or plt is None:
        missing = []
        if pd is None:
            missing.append("pandas")
        if plt is None:
            missing.append("matplotlib")
        print(f"⚠️ Skipping score analysis/plots because missing dependency: {', '.join(missing)}")
        print("   Install them if you need summary plots: pip install pandas matplotlib")
        return

    records = [
        json.loads(l)
        for l in scored_jsonl.read_text(encoding="utf-8").splitlines()
        if l.strip()
    ]

    rows, parse_fail_ids = [], []
    for rec in records:
        score = safe_float(rec.get("gemini_score"))
        if score is None:
            parse_fail_ids.append(rec.get("id"))
            continue
        rows.append(
            {
                "id": rec.get("id"),
                "model": rec.get("model_name"),
                "ability": rec.get("ability"),
                "score": score,
            }
        )

    print("len(rows):", len(rows))
    if not rows:
        print("❌ No usable score data, script terminated.")
        return

    df = pd.DataFrame(rows)

    BIG_CAT_W = 0.25
    WEIGHTS = {}

    ap_comp = "acoustic_attributes/composite_properties"
    ap_comp_w = BIG_CAT_W * 0.5
    ap_rem_each = (BIG_CAT_W - ap_comp_w) / (len(SUB_CATS["acoustic_attributes"]) - 1)

    for ab in SUB_CATS["acoustic_attributes"]:
        WEIGHTS[ab] = ap_comp_w if ab == ap_comp else ap_rem_each

    for big in ("instruction", "role_play", "empathy"):
        each = BIG_CAT_W / len(SUB_CATS[big])
        for ab in SUB_CATS[big]:
            WEIGHTS[ab] = each

    ordered_abilities = get_ordered_abilities()

    print("\n================= Statistical Results =================")
    for model, g in df.groupby("model"):
        print(f"\nModel: {model}")
        sub_means = g.groupby("ability")["score"].mean().to_dict()

        weighted_sum, used_w = 0.0, 0.0
        for ab, w in WEIGHTS.items():
            if ab in sub_means:
                weighted_sum += sub_means[ab] * w
                used_w += w
        weighted_mean = weighted_sum / used_w if used_w else float("nan")

        print(f"  Weighted overall average score: {weighted_mean:.2f} (weight coverage={used_w:.1%})")
        print("  Major category average scores:")

        for big_cat in ["acoustic_attributes", "instruction", "role_play", "empathy"]:
            big_cat_weighted_sum = 0.0
            big_cat_used_w = 0.0
            big_cat_sample_count = 0

            for ability in SUB_CATS[big_cat]:
                if ability in sub_means:
                    weight = WEIGHTS[ability]
                    big_cat_weighted_sum += sub_means[ability] * weight
                    big_cat_used_w += weight
                    big_cat_sample_count += g[g["ability"] == ability].shape[0]

            if big_cat_used_w > 0:
                big_cat_avg = big_cat_weighted_sum / big_cat_used_w
                coverage = big_cat_used_w / BIG_CAT_W
                print(f"    {big_cat:<20s}: {big_cat_avg:.2f} (n={big_cat_sample_count}, coverage={coverage:.1%})")
            else:
                print(f"    {big_cat:<20s}: N/A (no data)")

        print("  Ability average scores:")
        ability_means = g.groupby("ability")["score"].mean().to_dict()
        for ab in ordered_abilities:
            if ab in ability_means:
                cnt = g[g["ability"] == ab].shape[0]
                weight = WEIGHTS.get(ab, 0.0)
                print(f"    {ab:<40s}: {ability_means[ab]:.2f} (n={cnt}, weight={weight:.3f})")

    print("\nStarting visualization...")
    for model, g in df.groupby("model"):
        model_dir = out_dir / model
        ensure_dir(model_dir)

        sub_means = g.groupby("ability")["score"].mean().to_dict()
        big_cat_scores = []
        big_cat_labels = []

        for big_cat in ["acoustic_attributes", "instruction", "role_play", "empathy"]:
            big_cat_weighted_sum = 0.0
            big_cat_used_w = 0.0
            for ability in SUB_CATS[big_cat]:
                if ability in sub_means:
                    weight = WEIGHTS[ability]
                    big_cat_weighted_sum += sub_means[ability] * weight
                    big_cat_used_w += weight
            if big_cat_used_w > 0:
                big_cat_avg = big_cat_weighted_sum / big_cat_used_w
                big_cat_scores.append(big_cat_avg)
                big_cat_labels.append(big_cat.replace("_", " ").title())

        if big_cat_scores:
            plt.figure(figsize=(10, 6))
            bars = plt.bar(range(len(big_cat_scores)), big_cat_scores)
            plt.xticks(range(len(big_cat_labels)), big_cat_labels, rotation=45, ha="right")

            for bar, score in zip(bars, big_cat_scores):
                plt.text(
                    bar.get_x() + bar.get_width() / 2,
                    bar.get_height() + 0.01,
                    f"{score:.2f}",
                    ha="center",
                    va="bottom",
                )

            w_sum = sum(sub_means[ab] * w for ab, w in WEIGHTS.items() if ab in sub_means)
            w_used = sum(w for ab, w in WEIGHTS.items() if ab in sub_means)
            w_avg = w_sum / w_used if w_used else float("nan")

            plt.axhline(w_avg, linestyle="--", linewidth=2, label=f"Overall average: {w_avg:.2f}")
            plt.ylabel("Average Score")
            plt.title(f"{model} - Major Category Scores")
            plt.legend()
            plt.grid(True, alpha=0.3)
            plt.tight_layout()
            plt.savefig(model_dir / "major_category_scores.png", dpi=200, bbox_inches="tight")
            plt.close()

        ability_mean = g.groupby("ability")["score"].mean()
        ordered_data, ordered_labels = [], []
        for ab in ordered_abilities:
            if ab in ability_mean.index:
                ordered_data.append(ability_mean[ab])
                ordered_labels.append(ab)

        plt.figure(figsize=(8, 4 + 0.25 * len(ordered_data)))
        y_pos = range(len(ordered_data))
        plt.barh(y_pos, ordered_data)
        plt.yticks(y_pos, ordered_labels)

        sub_means = g.groupby("ability")["score"].mean().to_dict()
        w_sum = sum(sub_means[ab] * w for ab, w in WEIGHTS.items() if ab in sub_means)
        w_used = sum(w for ab, w in WEIGHTS.items() if ab in sub_means)
        w_avg = w_sum / w_used if w_used else float("nan")

        plt.axvline(w_avg, linestyle="--", label="Weighted overall average")
        plt.xlabel("Average score")
        plt.title(f"{model} - Ability average scores")
        plt.legend()
        plt.tight_layout()
        plt.savefig(model_dir / "ability_scores.png", dpi=200)
        plt.close()

    print(f"✅ Visualization complete, saved to folder: {out_dir}")

    print("\n================= Parse Failure Information =================")
    if parse_fail_ids:
        print(f"Total of {len(parse_fail_ids)} records failed gemini_score parsing.")
    else:
        print("All records' gemini_score parsed successfully.")


# ----------------------------- CLI ----------------------------------

def main():
    root = "./data/examples"

    parser = argparse.ArgumentParser(
        description="Concurrent proxy scoring for voice results with forced score retries"
    )
    parser.add_argument("--root_dir", default=f"{root}/model_res/en/wav", help="wav directory path")
    parser.add_argument("--metadata_path", default=f"{root}/model_res/en/metadata.jsonl", help="metadata.jsonl path")
    parser.add_argument("--prompts_dir", default="lalm_eval/eval_prompts/en", help="Directory containing prompt templates")
    parser.add_argument("--out_dir", default=f"{root}/eval_res/en", help="Output root directory")

    parser.add_argument("--temperature", type=float, default=1.0)
    parser.add_argument("--top_p", type=float, default=0.7)
    parser.add_argument("--max_tokens", type=int, default=4096)

    parser.add_argument(
        "--api_key",
        default=os.environ.get("VSTYLE_API_KEY") or os.environ.get("OPENAI_API_KEY"),
        help="Proxy API Key; defaults to VSTYLE_API_KEY or OPENAI_API_KEY",
    )
    parser.add_argument(
        "--base_url",
        default=os.environ.get("VSTYLE_BASE_URL"),
        help="Proxy base_url, e.g. https://xxx/v1; defaults to VSTYLE_BASE_URL",
    )
    parser.add_argument(
        "--model_name",
        default=os.environ.get("VSTYLE_JUDGE_MODEL", "gemini-2.5-pro"),
        help="Model name on proxy; defaults to VSTYLE_JUDGE_MODEL or gemini-2.5-pro",
    )

    parser.add_argument(
        "--max_retry_api",
        type=int,
        default=5,
        help="Maximum number of retry attempts, capped at 5 in scoring logic",
    )
    parser.add_argument(
        "--sleep_between_retry",
        type=int,
        default=5,
        help="Sleep time between retries (seconds)",
    )
    parser.add_argument(
        "--max_per_ability",
        type=int,
        default=100000,
        help="Maximum number of evaluations per ability",
    )
    parser.add_argument("--concurrency", type=int, default=4, help="Number of concurrent threads")
    parser.add_argument("--overwrite", action="store_true", help="Overwrite existing results")
    parser.add_argument(
        "--fallback_score",
        type=str,
        default="3",
        help="Fallback score if 5 attempts all fail; should be between 1 and 5",
    )
    parser.add_argument(
        "--dataset_metadata_path",
        type=str,
        default=None,
        help="原始 VStyle 数据集的 metadata.jsonl 路径，用于补充 instruct_text",
    )
    parser.add_argument("--verbose", action="store_true", help="Print raw judge replies during scoring")

    args = parser.parse_args()

    if not args.api_key:
        parser.error("--api_key is required, or set VSTYLE_API_KEY / OPENAI_API_KEY")
    if not args.base_url:
        parser.error("--base_url is required, or set VSTYLE_BASE_URL")
    if normalize_score_str(args.fallback_score) == "":
        raise ValueError("--fallback_score must be a number between 1 and 5")

    args.fallback_score = normalize_score_str(args.fallback_score)
    print(f"args: {args}")

    out_dir = Path(args.out_dir)
    ensure_dir(out_dir)

    scored_jsonl = eval_all_samples(
        root_dir=Path(args.root_dir),
        metadata_path=Path(args.metadata_path),
        prompts_dir=Path(args.prompts_dir),
        out_dir=out_dir,
        max_per_ability=args.max_per_ability,
        concurrency=args.concurrency,
        overwrite=args.overwrite,
        model_name=args.model_name,
        max_retry=args.max_retry_api,
        sleep_between_retry=args.sleep_between_retry,
        api_key=args.api_key,
        base_url=args.base_url,
        fallback_score=args.fallback_score,
        dataset_metadata_path=args.dataset_metadata_path,
        temperature=args.temperature,
        top_p=args.top_p,
        max_tokens=args.max_tokens,
        verbose=args.verbose,
    )

    analyze_scores(scored_jsonl, out_dir)


if __name__ == "__main__":
    main()