心理咨询师数字孪生对话数据集

自2023年5月发布SoulChat以来，我们经过对真实世界心理咨询语言风格、心理咨询技术等方面的深入探索，在心理咨询师数字孪生建模能力上取得了显著提升。ChatGPT诞生以来，国内外已有大量的工作将大模型（LLM）应用于情感陪护、心理健康支持对话、心理咨询对话领域，例如SoulChat、MeChat、QiaoBan、CPsyCoun、MindChat、EmoLLM等等。然而，过往的工作聚焦于借助精心设计的提示词来构建多轮心理健康对话数据集，微调出的“心理健康大模型”很容易造成回答的同质化、模板化，使得这些LLMs难以应对复杂多变的来访者，无法很好模拟现实世界真实心理咨询师的语言表达与疗法技术运用风格。针对上述问题，华南理工大学未来技术学院-广东省数字孪生人实验室在灵心大模型（SoulChat1.0）基础上，推出了心理咨询师数字孪生大模型SoulChat2.0。SoulChat2.0首次定义了特定心理咨询师的数字孪生（PsyDT, Psychological consultant Digital Twin）任务。为此，我们开源了9万多轮次高质量的心理咨询师数字孪生对话数据集PsyDTCorpus。它基于真实世界心理咨询师的少量真实咨询案例启发下构建得到。

Since the release of SoulChat in May 2023, we have conducted in-depth explorations on the linguistic styles and therapeutic techniques of real-world psychological counseling, and achieved remarkable improvements in the modeling capabilities of digital twins for psychological counselors. Since the advent of ChatGPT, a large number of domestic and international studies have applied large language models (LLMs) to the fields of emotional companionship, mental health support dialogue, and psychological counseling dialogue, such as SoulChat, MeChat, QiaoBan, CPsyCoun, MindChat, EmoLLM, etc. However, previous studies have focused on constructing multi-turn mental health dialogue datasets using carefully designed prompts. The fine-tuned "mental health LLMs" tend to produce homogeneous and formulaic responses, making these LLMs poorly equipped to handle complex and diverse clients, and unable to authentically mimic the linguistic expression and therapeutic technique application styles of real-world psychological counselors. To address these issues, the School of Future Technology, South China University of Technology and Guangdong Digital Twin Human Laboratory launched the psychological counselor digital twin LLM SoulChat 2.0 based on the Lingxin LLM (SoulChat 1.0). SoulChat 2.0 defines for the first time the task of digital twin for specific psychological counselors (PsyDT, Psychological Consultant Digital Twin). To this end, we have open-sourced PsyDTCorpus, a high-quality psychological counselor digital twin dialogue dataset with over 90,000 turns, which was constructed based on insights from a small number of real counseling cases from real-world psychological counselors.

搜集汇总

背景与挑战

背景概述

该数据集是心理咨询师数字孪生对话数据集，旨在通过少量真实咨询案例生成大规模高质量对话数据，以捕捉特定心理咨询师的语言风格和疗法技术。数据集包含5000个对话共90,365轮，分为训练集和测试集，采用OpenAI格式，专注于心理健康话题，用于微调大语言模型以提升心理咨询性能。