tiptoghosh/food-recipes-15k

Food.com多模态食谱数据集（15K）是一个精心策划的高质量多模态数据集，专为训练和评估视觉语言模型（VLMs）和多模态检索增强生成（RAG）系统而设计。每个样本包含一个渲染的食谱卡片图像（PNG格式，300 DPI）和同一食谱的结构化Markdown表示，支持跨模态检索、VLM微调和文档理解研究。数据集来源于Food.com Recipes and User Interactions（Kaggle），经过六阶段的可重复管道处理，最终筛选出约15,000个样本。数据集支持多种下游任务，如多模态RAG、VLM微调、跨模态检索和文档布局理解。

The Food.com Multimodal Recipe Dataset (15K) is a curated, high-quality multimodal recipe dataset engineered for training and evaluating vision-language models (VLMs) and multimodal retrieval-augmented generation (RAG) systems. Each sample pairs a rendered recipe card image (PNG, 300 DPI) with a structured Markdown representation of the same recipe, enabling cross-modal retrieval, VLM fine-tuning, and document understanding research. The dataset is sourced from Food.com Recipes and User Interactions (Kaggle) and processed through a six-stage reproducible pipeline, resulting in approximately 15,000 samples. It supports various downstream tasks such as multimodal RAG, VLM fine-tuning, cross-modal retrieval, and document layout understanding.