宜博 / lemonboss

#!/usr/bin/env python3
"""Benchmark 评估器 - create-drama 技能"""

import json
import sys
import re

def evaluate(result_json, benchmark_file, testdata_dir):
    """评估剧本生成结果"""

    try:
        result = json.loads(result_json)
    except:
        return {
            "score": 50,
            "dimensions": {},
            "reason": "无法解析输出JSON"
        }

    scores = {}

    # D1: 素材分析准确性 (20%)
    # 检查是否识别素材类型、提取核心元素、语言检测、风格适配
    name = result.get('name', '')
    title = result.get('title', '')
    description = result.get('description', '')
    language = result.get('language', 'zh')

    # 核心元素评估
    d1_score = 0

    # 角色和冲突提取 - 检查剧本是否有角色和台词(代表有故事)
    characters = result.get('characters', [])
    script = result.get('script', [])

    if len(characters) >= 2:
        d1_score += 25
    if len(script) >= 5:
        d1_score += 25

    # 语言检测 - 中文字符 vs 英文字符
    has_cn = any('\u4e00' <= c <= '\u9fff' for c in (name + title + description))
    if (language == 'zh' and has_cn) or (language == 'en' and not has_cn):
        d1_score += 25
    elif language in ['zh', 'en', 'multi']:
        d1_score += 15

    # 风格适配检测
    description_len = len(description)
    if description_len >= 10:
        d1_score += 25

    scores['D1'] = min(100, d1_score)

    # D2: 角色配置合理性 (20%)
    # 角色数量(2-5)、音色匹配、站位配置
    d2_score = 0

    char_count = len(characters)
    if 2 <= char_count <= 5:
        d2_score += 40
    elif char_count > 5:
        d2_score += 20

    # 音色匹配检查
    valid_voices = ['zh_male_m191', 'zh_female_cancan', 'zh_male_shaonianzixin',
                   'zh_female_sajiaoxuemei', 'zh_male_dayi', 'zh_male_sunwukong',
                   'en_male_tim', 'en_female_dacey', 'en_female_stokie']
    voice_match = 0
    for char in characters:
        voice = char.get('voiceId', '')
        if any(v in voice for v in valid_voices):
            voice_match += 1
    if char_count > 0:
        voice_ratio = voice_match / char_count
        d2_score += int(60 * voice_ratio)

    scores['D2'] = min(100, d2_score)

    # D3: 剧本结构完整性 (20%)
    # Classic 结构(开场2→铺垫4→冲突6→反转4→结局2)、台词数量(10-20)、台词≤30字
    d3_score = 0

    # 台词数量
    script_count = len(script)
    if 10 <= script_count <= 20:
        d3_score += 30
    elif 5 <= script_count < 10:
        d3_score += 20
    elif script_count > 0:
        d3_score += 10

    # 台词长度检查
    long_lines = 0
    for line in script:
        dialogue = line.get('dialogue', '')
        if len(dialogue) > 30:
            long_lines += 1

    long_ratio = long_lines / script_count if script_count > 0 else 1
    if long_ratio < 0.2:
        d3_score += 30
    elif long_ratio < 0.5:
        d3_score += 20
    else:
        d3_score += 10

    # 结构完整性 - 检查是否有 position 字段
    has_position = any('position' in line for line in script)
    if has_position:
        d3_score += 40
    else:
        d3_score += 20

    scores['D3'] = min(100, d3_score)

    # D4: 图片生成质量 (20%)
    # 立绘 prompt 含空白脸部、背景图/封面图 prompt
    d4_score = 0

    # 检查空白脸部
    blank_face_count = 0
    for char in characters:
        prompt = char.get('imagePrompt', '')
        if '空白' in prompt or '不要画' in prompt or '纯肉色' in prompt or 'blank' in prompt.lower():
            blank_face_count += 1

    if char_count > 0:
        blank_ratio = blank_face_count / char_count
        if blank_ratio >= 0.8:
            d4_score += 60
        elif blank_ratio >= 0.5:
            d4_score += 40
        elif blank_ratio > 0:
            d4_score += 20

    # 背景图
    background_prompt = result.get('backgroundPrompt', '')
    if background_prompt:
        d4_score += 20
    else:
        d4_score += 10

    # 封面图
    cover_prompt = result.get('coverPrompt', '')
    if cover_prompt:
        d4_score += 20
    else:
        d4_score += 10

    scores['D4'] = min(100, d4_score)

    # D5: 数据库写入正确性 (20%)
    # JSON 格式规范性、必填字段完整性
    d5_score = 0

    # 必填字段检查
    required_fields = ['name', 'title', 'language', 'characters', 'script']
    field_count = sum(1 for f in required_fields if f in result and result[f])
    d5_score += int(40 * (field_count / len(required_fields)))

    # systemPrompt 检查
    system_prompt = result.get('systemPrompt', '')
    if len(system_prompt) >= 100:
        d5_score += 30
    elif len(system_prompt) >= 50:
        d5_score += 20
    elif system_prompt:
        d5_score += 10

    # tags 和 description 检查
    tags = result.get('tags', '')
    if tags and description:
        d5_score += 30
    elif tags or description:
        d5_score += 15

    scores['D5'] = min(100, d5_score)

    # 计算总分
    total = (scores['D1'] * 0.20 + scores['D2'] * 0.20 +
             scores['D3'] * 0.20 + scores['D4'] * 0.20 +
             scores['D5'] * 0.20)

    # 评分等级
    if total >= 95:
        grade = 'A+'
    elif total >= 85:
        grade = 'A'
    elif total >= 80:
        grade = 'B+'
    elif total >= 70:
        grade = 'B'
    elif total >= 60:
        grade = 'C'
    else:
        grade = 'D'

    result_output = {
        "score": round(total, 2),
        "grade": grade,
        "dimensions": scores,
        "breakdown": {
            "D1_素材分析准确性": f"{scores['D1']} × 20% = {scores['D1'] * 0.20:.1f}",
            "D2_角色配置合理性": f"{scores['D2']} × 20% = {scores['D2'] * 0.20:.1f}",
            "D3_剧本结构完整性": f"{scores['D3']} × 20% = {scores['D3'] * 0.20:.1f}",
            "D4_图片生成质量": f"{scores['D4']} × 20% = {scores['D4'] * 0.20:.1f}",
            "D5_数据库写入正确性": f"{scores['D5']} × 20% = {scores['D5'] * 0.20:.1f}"
        },
        "reason": "基于 benchmark_v1.md 5维度评估",
        "char_count": char_count,
        "script_count": script_count
    }

    print(json.dumps(result_output, indent=2, ensure_ascii=False))
    return result_output

if __name__ == '__main__':
    result_file = '/tmp/result.json'
    with open(result_file, 'r') as f:
        result_json = f.read()
    evaluate(result_json, '', '')