mrdbourke/recipe-synthetic-images-10k

recipe-synthetic-images-10k is designed to practice multimodal RAG (Retrieval Augmented Generation).

The initial recipes were sourced from the Food.com dataset on Kaggle.

A sample of 10k recipes was taken from the dataset (the original total was 200k+) with the following criteria:

1. 1.n_ingredients ranging from 1-20 (max 20 ingredients).
2. 2.n_reviews >= 5 (minimum 5 reviews).

Each recipe was then turned into a PDF and rendered to an image using two Python scripts provided by Claude with some guidance.

The scripts were designed to be simple enough to have a workflow of:

Recipe JSON input -> Random layout selection (keeping to 1 page) -> Output PDF + image

This dataset can now be used in a few ways:

1. 1.Practice text-only RAG - Use the text-based recipes to feed into a RAG pipeline.
2. 2.Practice image-only RAG - Use the image-based recipes to feed into a RAG pipeline.
3. 3.Practice image + text RAG - Use both the image and text-based recipes to feed into a RAG pipeline.

Source

This dataset was created live on YouTube by Daniel Bourke.