arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2407.00102cs.LGcs.AIcs.CL
收藏

基于质量驱动数据选择的课程学习

Curriculum Learning with Quality-Driven Data Selection

Australian Artificial Intelligence Institute(澳大利亚人工智能研究所)

Biao Wu, Ling Chen

AI总结:

本文提出一种基于图像-文本相关性和模型困惑度的二维质量空间数据选择方法,用于构建多阶段课程学习子集,从而提升多模态大语言模型的视觉指令微调效果。

摘要:

OpenAI 的 GPT-4 所展示的令人印象深刻的多模态能力,引发了人们对多模态大语言模型(MLLMs)开发的浓厚兴趣。使用机器生成的指令遵循数据对 MLLMs 进行视觉指令微调,已被证明能够增强其在各种任务上的零样本能力。然而,目前对指令数据质量控制的探索仍然有限。当前 MLLMs 中的数据选择方法通常依赖单一且不可靠的分数,或使用下游任务进行选择,这不仅耗时,还可能导致在所选的评估数据集上出现过拟合。为了缓解这些局限性,我们提出了一种新的数据选择方法,该方法利用图像-文本相关性和模型困惑度来评估并选择不同质量的数据。该方法利用这两个属性的独特分布,将数据质量映射到一个二维空间中,从而可以根据数据在该分布中的位置进行选择。通过利用这一空间,我们可以分析作为提示使用的任务类型设置对数据质量的影响。此外,该空间还可用于构建不同质量的多阶段子集,以促进课程学习。我们的研究包括在多个数据集上进行的全面实验。结果表明,与使用完整数据集相比,在五种常用评估能力上均取得了显著提升。我们的代码、数据和模型已公开,地址为:https://anonymous.4open.science/r/EHIT-31B4

英文摘要:

The impressive multimodal capabilities demonstrated by OpenAI's GPT-4 have generated significant interest in the development of Multimodal Large Language Models (MLLMs). Visual instruction tuning of MLLMs with machine-generated instruction-following data has shown to enhance zero-shot capabilities across various tasks. However, there has been limited exploration into controlling the quality of the instruction data.Current methodologies for data selection in MLLMs often rely on single, unreliable scores or use downstream tasks for selection, which is time-consuming and can lead to potential overfitting on the chosen evaluation datasets. To mitigate these limitations, we propose a novel data selection methodology that utilizes image-text correlation and model perplexity to evaluate and select data of varying quality. This approach leverages the distinct distribution of these two attributes, mapping data quality into a two-dimensional space that allows for the selection of data based on their location within this distribution. By utilizing this space, we can analyze the impact of task type settings, used as prompts, on data quality. Additionally, this space can be used to construct multi-stage subsets of varying quality to facilitate curriculum learning. Our research includes comprehensive experiments conducted on various datasets. The results emphasize substantial enhancements in five commonly assessed capabilities compared to using the complete dataset. Our codes, data, and models are publicly available at: https://anonymous.4open.science/r/EHIT-31B4

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑