数据集获取

git clone https://github.com/lemonsis/MDD-5k.git

curl -L -o repo.zip https://github.com/lemonsis/MDD-5k/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

MDD-5k

Official Implementation of AAAI 2025 paper “MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents”

Dataset Overview

MDD-5k is build upon 926 unique patient files, and each patient file is used to synthesize five diagnostic conversations, resulting in a total of 4630 conversations.
The synthetic conversations and corresponding lables can be found in ./MDD_5k and ./Label. Each json file in ./MDD_5k contains five conversations and all the five conversation is generated based on the same patient file.

Preprocessing Steps

Run python patient_template_gen.py to get statistics of MDD-5k dataset and generate fictitious patient experiences

Synthesize Diagnostic Conversation

Current code only supports using deployed model for generation. You can either access state-of-the-art GPT model through OpenAI API keys, or deploying local models with vLLM. For the former way, enter your OpenAI key in line 37 of llm_tools_api.py. For the latter way, enter the key and server host in line 44 and 45 of llm_tools_api.py. For OpenAI LLMs, we found the performance of gpt-4o is the best.
Then run python main.py to start synthesizing data.

Example

One real patient case is shown in ./raw_data/pa20.json. Five fictitious patient experiences generated with this patient case are shown in ./prompts/background_story/patient_1. The synthesized diagnostic conversations are in ./DataSyn/patient1.json.

数据加载示例（表格/文本类）

import pandas as pd, glob, os

files = (glob.glob(os.path.join(path, "**", "*.csv"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tsv"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.xlsx"), recursive=True))
print("数据文件:", files)
df = pd.read_csv(files[0])
print(df.shape); print(df.columns.tolist()); print(df.head(3))

完整仓库：github.com/lemonsis/MDD-5k