WaltonFuture/InstructionGPT-4
收藏资源简介:
--- task_categories: - visual-question-answering size_categories: - n<1K --- # InstructionGPT-4: A 200-Instruction Paradigm for Fine-Tuning MiniGPT-4 [Lai Wei](https://waltonfuture.github.io/), Zihao Jiang, [Weiran Huang](https://www.weiranhuang.com/), [Lichao Sun](https://lichao-sun.github.io/). **Shanghai Jiao Tong University, Lehigh University** [Paper](https://arxiv.org/abs/2308.12067), [Link](https://mp.weixin.qq.com/s/s4Acec71v5oMlFkyhlCL_g), [Code](https://github.com/waltonfuture/InstructionGPT-4) ## Introduction Multimodal large language models acquire their instruction-following capabilities through a two-stage training process: pre-training on image-text pairs and fine-tuning on supervised vision-language instruction data. Recent studies have shown that large language models can achieve satisfactory results even with a limited amount of high-quality instruction-following data. In this paper, we introduce InstructionGPT-4, which is fine-tuned on a small dataset comprising only 200 examples, amounting to approximately 6% of the instruction-following data used in the alignment dataset for MiniGPT-4. We first propose several metrics to access the quality of multimodal instruction data. Based on these metrics, we present a simple and effective data selector to automatically identify and filter low-quality vision-language data. By employing this method, InstructionGPT-4 outperforms the original MiniGPT-4 on various evaluations (e.g., visual question answering, GPT-4 preference). Overall, our findings demonstrate that less but high-quality instruction tuning data is efficient to enable multimodal large language models to generate better output. ## Usage You can download our vision-language dataset containing only 200 high-quality examples and replace the original cc_sbu_align dataset used in the fine-tuning stage of MiniGPT-4. The training settings are the same as [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4). If you're using InstructionGPT-4 in your research or applications, please cite using this BibTeX: ```bibtex @article{wei2023instructiongpt, title={InstructionGPT-4: A 200-Instruction Paradigm for Fine-Tuning MiniGPT-4}, author={Wei, Lai and Jiang, Zihao and Huang, Weiran and Sun, Lichao}, journal={arXiv preprint arXiv:2308.12067}, year={2023} } ```
任务类别: - 视觉问答(visual-question-answering) 样本量类别: - n<1K # InstructionGPT-4:面向MiniGPT-4微调的200条指令范式 [赖伟(Lai Wei)](https://waltonfuture.github.io/), 江梓豪, [黄伟然(Weiran Huang)](https://www.weiranhuang.com/), [孙立超(Lichao Sun)](https://lichao-sun.github.io/). **上海交通大学,利哈伊大学** [论文](https://arxiv.org/abs/2308.12067), [微信链接](https://mp.weixin.qq.com/s/s4Acec71v5oMlFkyhlCL_g), [代码](https://github.com/waltonfuture/InstructionGPT-4) ## 引言 多模态大语言模型通过两阶段训练流程获得指令遵循能力:先在图像-文本配对数据上进行预训练,再在监督式视觉语言指令数据上进行微调。近期研究表明,即便仅使用少量高质量的指令遵循数据,大语言模型也能取得令人满意的效果。本文提出InstructionGPT-4,其仅基于包含200条样本的小型数据集完成微调,该数据集规模约为MiniGPT-4对齐阶段所用指令遵循数据的6%。本文首先提出若干指标以评估多模态指令数据的质量,并基于这些指标设计了一款简单高效的数据选择器,可自动识别并过滤低质量的视觉语言数据。通过应用该方法,InstructionGPT-4在多项评估任务(如视觉问答、GPT-4偏好性评估)中均优于原始MiniGPT-4。综上,本研究结果证实,使用更少但高质量的指令微调数据,能够有效赋能多模态大语言模型生成更优质的输出。 ## 使用方法 您可下载仅包含200条高质量样本的视觉语言数据集,替换MiniGPT-4微调阶段所使用的原始cc_sbu_align数据集。训练设置与[MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4)保持一致。 若您在研究或应用中使用InstructionGPT-4,请引用如下BibTeX格式的条目: bibtex @article{wei2023instructiongpt, title={InstructionGPT-4: A 200-Instruction Paradigm for Fine-Tuning MiniGPT-4}, author={Wei, Lai and Jiang, Zihao and Huang, Weiran and Sun, Lichao}, journal={arXiv preprint arXiv:2308.12067}, year={2023} }
InstructionGPT-4 数据集概述
基本信息
- 任务类别: 视觉问答 (visual-question-answering)
- 数据集规模: 小于1000条 (n<1K)
数据集介绍
- 名称: InstructionGPT-4
- 来源: 上海交通大学, 理海大学
- 作者: Lai Wei, Zihao Jiang, Weiran Huang, Lichao Sun
- 相关链接: 论文, 代码
数据集详情
- 数据集大小: 仅包含200个高质量示例
- 用途: 用于替换MiniGPT-4微调阶段的原始cc_sbu_align数据集
- 训练设置: 与MiniGPT-4相同
引用信息
bibtex @article{wei2023instructiongpt, title={InstructionGPT-4: A 200-Instruction Paradigm for Fine-Tuning MiniGPT-4}, author={Wei, Lai and Jiang, Zihao and Huang, Weiran and Sun, Lichao}, journal={arXiv preprint arXiv:2308.12067}, year={2023} }




