308 lines
12 KiB
Python
308 lines
12 KiB
Python
from typing import Dict, List
|
||
import json
|
||
import os
|
||
from dotenv import load_dotenv
|
||
import sys
|
||
import toml
|
||
import random
|
||
from tqdm import tqdm
|
||
|
||
# 添加项目根目录到 Python 路径
|
||
root_path = os.path.abspath(os.path.join(os.path.dirname(__file__), "../.."))
|
||
sys.path.append(root_path)
|
||
|
||
# 加载配置文件
|
||
config_path = os.path.join(root_path, "config", "bot_config.toml")
|
||
with open(config_path, "r", encoding="utf-8") as f:
|
||
config = toml.load(f)
|
||
|
||
# 现在可以导入src模块
|
||
from src.individuality.scene import get_scene_by_factor, PERSONALITY_SCENES #noqa E402
|
||
from src.individuality.questionnaire import FACTOR_DESCRIPTIONS #noqa E402
|
||
from src.individuality.offline_llm import LLM_request_off #noqa E402
|
||
|
||
# 加载环境变量
|
||
env_path = os.path.join(root_path, ".env")
|
||
if os.path.exists(env_path):
|
||
print(f"从 {env_path} 加载环境变量")
|
||
load_dotenv(env_path)
|
||
else:
|
||
print(f"未找到环境变量文件: {env_path}")
|
||
print("将使用默认配置")
|
||
|
||
|
||
def adapt_scene(scene: str) -> str:
|
||
|
||
personality_core = config['personality']['personality_core']
|
||
personality_sides = config['personality']['personality_sides']
|
||
personality_side = random.choice(personality_sides)
|
||
identity_details = config['identity']['identity_detail']
|
||
identity_detail = random.choice(identity_details)
|
||
|
||
"""
|
||
根据config中的属性，改编场景使其更适合当前角色
|
||
|
||
Args:
|
||
scene: 原始场景描述
|
||
|
||
Returns:
|
||
str: 改编后的场景描述
|
||
"""
|
||
try:
|
||
prompt = f"""
|
||
这是一个参与人格测评的角色形象:
|
||
- 昵称: {config['bot']['nickname']}
|
||
- 性别: {config['identity']['gender']}
|
||
- 年龄: {config['identity']['age']}岁
|
||
- 外貌: {config['identity']['appearance']}
|
||
- 性格核心: {personality_core}
|
||
- 性格侧面: {personality_side}
|
||
- 身份细节: {identity_detail}
|
||
|
||
请根据上述形象，改编以下场景，在测评中，用户将根据该场景给出上述角色形象的反应:
|
||
{scene}
|
||
保持场景的本质不变，但最好贴近生活且具体，并且让它更适合这个角色。
|
||
改编后的场景应该自然、连贯，并考虑角色的年龄、身份和性格特点。只返回改编后的场景描述，不要包含其他说明。注意{config['bot']['nickname']}是面对这个场景的人，而不是场景的其他人。场景中不会有其描述，
|
||
现在，请你给出改编后的场景描述
|
||
"""
|
||
|
||
llm = LLM_request_off(model_name=config['model']['llm_normal']['name'])
|
||
adapted_scene, _ = llm.generate_response(prompt)
|
||
|
||
# 检查返回的场景是否为空或错误信息
|
||
if not adapted_scene or "错误" in adapted_scene or "失败" in adapted_scene:
|
||
print("场景改编失败，将使用原始场景")
|
||
return scene
|
||
|
||
return adapted_scene
|
||
except Exception as e:
|
||
print(f"场景改编过程出错：{str(e)}，将使用原始场景")
|
||
return scene
|
||
|
||
|
||
class PersonalityEvaluator_direct:
|
||
def __init__(self):
|
||
self.personality_traits = {"开放性": 0, "严谨性": 0, "外向性": 0, "宜人性": 0, "神经质": 0}
|
||
self.scenarios = []
|
||
self.final_scores = {"开放性": 0, "严谨性": 0, "外向性": 0, "宜人性": 0, "神经质": 0}
|
||
self.dimension_counts = {trait: 0 for trait in self.final_scores.keys()}
|
||
|
||
# 为每个人格特质获取对应的场景
|
||
for trait in PERSONALITY_SCENES:
|
||
scenes = get_scene_by_factor(trait)
|
||
if not scenes:
|
||
continue
|
||
|
||
# 从每个维度选择3个场景
|
||
import random
|
||
|
||
scene_keys = list(scenes.keys())
|
||
selected_scenes = random.sample(scene_keys, min(3, len(scene_keys)))
|
||
|
||
for scene_key in selected_scenes:
|
||
scene = scenes[scene_key]
|
||
|
||
# 为每个场景添加评估维度
|
||
# 主维度是当前特质，次维度随机选择一个其他特质
|
||
other_traits = [t for t in PERSONALITY_SCENES if t != trait]
|
||
secondary_trait = random.choice(other_traits)
|
||
|
||
self.scenarios.append(
|
||
{"场景": scene["scenario"], "评估维度": [trait, secondary_trait], "场景编号": scene_key}
|
||
)
|
||
|
||
self.llm = LLM_request_off()
|
||
|
||
def evaluate_response(self, scenario: str, response: str, dimensions: List[str]) -> Dict[str, float]:
|
||
"""
|
||
使用 DeepSeek AI 评估用户对特定场景的反应
|
||
"""
|
||
# 构建维度描述
|
||
dimension_descriptions = []
|
||
for dim in dimensions:
|
||
desc = FACTOR_DESCRIPTIONS.get(dim, "")
|
||
if desc:
|
||
dimension_descriptions.append(f"- {dim}：{desc}")
|
||
|
||
dimensions_text = "\n".join(dimension_descriptions)
|
||
|
||
prompt = f"""请根据以下场景和用户描述，评估用户在大五人格模型中的相关维度得分（1-6分）。
|
||
|
||
场景描述：
|
||
{scenario}
|
||
|
||
用户回应：
|
||
{response}
|
||
|
||
需要评估的维度说明：
|
||
{dimensions_text}
|
||
|
||
请按照以下格式输出评估结果（仅输出JSON格式）：
|
||
{{
|
||
"{dimensions[0]}": 分数,
|
||
"{dimensions[1]}": 分数
|
||
}}
|
||
|
||
评分标准：
|
||
1 = 非常不符合该维度特征
|
||
2 = 比较不符合该维度特征
|
||
3 = 有点不符合该维度特征
|
||
4 = 有点符合该维度特征
|
||
5 = 比较符合该维度特征
|
||
6 = 非常符合该维度特征
|
||
|
||
请根据用户的回应，结合场景和维度说明进行评分。确保分数在1-6之间，并给出合理的评估。"""
|
||
|
||
try:
|
||
ai_response, _ = self.llm.generate_response(prompt)
|
||
# 尝试从AI响应中提取JSON部分
|
||
start_idx = ai_response.find("{")
|
||
end_idx = ai_response.rfind("}") + 1
|
||
if start_idx != -1 and end_idx != 0:
|
||
json_str = ai_response[start_idx:end_idx]
|
||
scores = json.loads(json_str)
|
||
# 确保所有分数在1-6之间
|
||
return {k: max(1, min(6, float(v))) for k, v in scores.items()}
|
||
else:
|
||
print("AI响应格式不正确，使用默认评分")
|
||
return {dim: 3.5 for dim in dimensions}
|
||
except Exception as e:
|
||
print(f"评估过程出错：{str(e)}")
|
||
return {dim: 3.5 for dim in dimensions}
|
||
|
||
def run_evaluation(self):
|
||
"""
|
||
运行整个评估过程
|
||
"""
|
||
print(f"欢迎使用{config['bot']['nickname']}形象创建程序！")
|
||
print("接下来，将给您呈现一系列有关您bot的场景（共15个）。")
|
||
print("请想象您的bot在以下场景下会做什么，并描述您的bot的反应。")
|
||
print("每个场景都会进行不同方面的评估。")
|
||
print("\n角色基本信息：")
|
||
print(f"- 昵称：{config['bot']['nickname']}")
|
||
print(f"- 性格核心：{config['personality']['personality_core']}")
|
||
print(f"- 性格侧面：{config['personality']['personality_sides']}")
|
||
print(f"- 身份细节：{config['identity']['identity_detail']}")
|
||
print("\n准备好了吗？按回车键开始...")
|
||
input()
|
||
|
||
total_scenarios = len(self.scenarios)
|
||
progress_bar = tqdm(total=total_scenarios, desc="场景进度", ncols=100, bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}]')
|
||
|
||
for _i, scenario_data in enumerate(self.scenarios, 1):
|
||
# print(f"\n{'-' * 20} 场景 {i}/{total_scenarios} - {scenario_data['场景编号']} {'-' * 20}")
|
||
|
||
# 改编场景，使其更适合当前角色
|
||
print(f"{config['bot']['nickname']}祈祷中...")
|
||
adapted_scene = adapt_scene(scenario_data["场景"])
|
||
scenario_data["改编场景"] = adapted_scene
|
||
|
||
print(adapted_scene)
|
||
print(f"\n请描述{config['bot']['nickname']}在这种情况下会如何反应：")
|
||
response = input().strip()
|
||
|
||
if not response:
|
||
print("反应描述不能为空！")
|
||
continue
|
||
|
||
print("\n正在评估您的描述...")
|
||
scores = self.evaluate_response(adapted_scene, response, scenario_data["评估维度"])
|
||
|
||
# 更新最终分数
|
||
for dimension, score in scores.items():
|
||
self.final_scores[dimension] += score
|
||
self.dimension_counts[dimension] += 1
|
||
|
||
print("\n当前评估结果：")
|
||
print("-" * 30)
|
||
for dimension, score in scores.items():
|
||
print(f"{dimension}: {score}/6")
|
||
|
||
# 更新进度条
|
||
progress_bar.update(1)
|
||
|
||
# if i < total_scenarios:
|
||
# print("\n按回车键继续下一个场景...")
|
||
# input()
|
||
|
||
progress_bar.close()
|
||
|
||
# 计算平均分
|
||
for dimension in self.final_scores:
|
||
if self.dimension_counts[dimension] > 0:
|
||
self.final_scores[dimension] = round(self.final_scores[dimension] / self.dimension_counts[dimension], 2)
|
||
|
||
print("\n" + "=" * 50)
|
||
print(f" {config['bot']['nickname']}的人格特征评估结果 ".center(50))
|
||
print("=" * 50)
|
||
for trait, score in self.final_scores.items():
|
||
print(f"{trait}: {score}/6".ljust(20) + f"测试场景数：{self.dimension_counts[trait]}".rjust(30))
|
||
print("=" * 50)
|
||
|
||
# 返回评估结果
|
||
return self.get_result()
|
||
|
||
def get_result(self):
|
||
"""
|
||
获取评估结果
|
||
"""
|
||
return {
|
||
"final_scores": self.final_scores,
|
||
"dimension_counts": self.dimension_counts,
|
||
"scenarios": self.scenarios,
|
||
"bot_info": {
|
||
"nickname": config['bot']['nickname'],
|
||
"gender": config['identity']['gender'],
|
||
"age": config['identity']['age'],
|
||
"height": config['identity']['height'],
|
||
"weight": config['identity']['weight'],
|
||
"appearance": config['identity']['appearance'],
|
||
"personality_core": config['personality']['personality_core'],
|
||
"personality_sides": config['personality']['personality_sides'],
|
||
"identity_detail": config['identity']['identity_detail']
|
||
}
|
||
}
|
||
|
||
|
||
def main():
|
||
evaluator = PersonalityEvaluator_direct()
|
||
result = evaluator.run_evaluation()
|
||
|
||
# 准备简化的结果数据
|
||
simplified_result = {
|
||
"openness": round(result["final_scores"]["开放性"] / 6, 1), # 转换为0-1范围
|
||
"conscientiousness": round(result["final_scores"]["严谨性"] / 6, 1),
|
||
"extraversion": round(result["final_scores"]["外向性"] / 6, 1),
|
||
"agreeableness": round(result["final_scores"]["宜人性"] / 6, 1),
|
||
"neuroticism": round(result["final_scores"]["神经质"] / 6, 1),
|
||
"bot_nickname": config['bot']['nickname']
|
||
}
|
||
|
||
# 确保目录存在
|
||
save_dir = os.path.join(root_path, "data", "personality")
|
||
os.makedirs(save_dir, exist_ok=True)
|
||
|
||
# 创建文件名，替换可能的非法字符
|
||
bot_name = config['bot']['nickname']
|
||
# 替换Windows文件名中不允许的字符
|
||
for char in ['\\', '/', ':', '*', '?', '"', '<', '>', '|']:
|
||
bot_name = bot_name.replace(char, '_')
|
||
|
||
file_name = f"{bot_name}_personality.per"
|
||
save_path = os.path.join(save_dir, file_name)
|
||
|
||
# 保存简化的结果
|
||
with open(save_path, "w", encoding="utf-8") as f:
|
||
json.dump(simplified_result, f, ensure_ascii=False, indent=4)
|
||
|
||
print(f"\n结果已保存到 {save_path}")
|
||
|
||
# 同时保存完整结果到results目录
|
||
os.makedirs("results", exist_ok=True)
|
||
with open("results/personality_result.json", "w", encoding="utf-8") as f:
|
||
json.dump(result, f, ensure_ascii=False, indent=2)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|