Rationale-Augmented Dataset
收藏资源简介:
Rationale-Augmented Dataset是一个包含指令和理由对的数据集,旨在帮助LVLMs模型在解释相关图像时生成更加准确的文本内容,减少幻觉现象。数据集的创建过程涉及使用LLM模型生成理由说明,并将这些理由嵌入到原始指令中。数据集的应用领域主要是多模态推理任务,如图像描述、视觉问答和多模态对话。通过使用Rationale-Augmented Dataset进行微调,模型在幻觉特定任务和更广泛的多模态推理任务上都有显著提升。
The Rationale-Augmented Dataset is a dataset comprising instruction-rationale pairs, designed to help large vision-language models (LVLMs) generate more accurate textual content when interpreting relevant images and reduce hallucinations. The dataset is constructed by generating rationales via large language models (LLMs) and embedding these rationales into the original instructions. It is primarily applied to multimodal reasoning tasks, such as image captioning, visual question answering (VQA), and multimodal dialogue. Fine-tuning models using the Rationale-Augmented Dataset leads to significant improvements in both hallucination-specific tasks and broader multimodal reasoning tasks.

- 1Critique Before Thinking: Mitigating Hallucination through Rationale-Augmented Instruction Tuning中国科学院信息工程研究所, 腾讯PCG基础技术中心 · 2025年



