Scholarlink

This research introduces LLaVA-Docent-V2, an enhanced version of the LLaVA-Docent model, which utilizes the capabilities of Large Multimodal Models (LMM) in art appreciation education. Building on its predecessor, LLaVA-Docent-V2 incorporates superior data structures, significantly improving the quality and scope of its educational offerings. An exhaustive review informed the improvements in relevant literature data and in-depth consultations with subject matter experts, contributing to a richer, more detailed training dataset featuring advanced virtual dialogues. A detailed comparative analysis of LLaVA-Docent-V2 against its precursor was conducted, involving both quantitative metrics and qualitative evaluations by a panel of expert researchers. The results affirm the substantial enhancements in LLaVA-Docent-V2, particularly its effectiveness in delivering comprehensive art education. The model provides exhaustive content across various critical stages, empowers students to delve into topics of interest, and proficiently identifies and corrects mistakes. However, LLaVA-Docent-V2 occasionally presents factual errors about artwork information and needs help with personalization. Addressing these issues will be crucial in the model’s ongoing development and refinement.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLaVA-Docent-V2: Improving Data Quality and Pedagogical Data Generation to Train Large Multimodal Models for Art Appreciation Education

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Unggi Lee,
  • Yoorim Son,
  • Jaeyoon Shin,
  • Gyuri Byun,
  • Yunseo Lee,
  • Junbo Koh,
  • Minji Jeon,
  • Hyeoncheol Kim

摘要

This research introduces LLaVA-Docent-V2, an enhanced version of the LLaVA-Docent model, which utilizes the capabilities of Large Multimodal Models (LMM) in art appreciation education. Building on its predecessor, LLaVA-Docent-V2 incorporates superior data structures, significantly improving the quality and scope of its educational offerings. An exhaustive review informed the improvements in relevant literature data and in-depth consultations with subject matter experts, contributing to a richer, more detailed training dataset featuring advanced virtual dialogues. A detailed comparative analysis of LLaVA-Docent-V2 against its precursor was conducted, involving both quantitative metrics and qualitative evaluations by a panel of expert researchers. The results affirm the substantial enhancements in LLaVA-Docent-V2, particularly its effectiveness in delivering comprehensive art education. The model provides exhaustive content across various critical stages, empowers students to delve into topics of interest, and proficiently identifies and corrects mistakes. However, LLaVA-Docent-V2 occasionally presents factual errors about artwork information and needs help with personalization. Addressing these issues will be crucial in the model’s ongoing development and refinement.