Recipe1M+
MIT CSAIL and collaborators
About a million scraped cooking recipes paired with millions of food photos, for research use only.
Recipe1M+ pairs recipes scraped from cooking websites (title, ingredient list, and step-by-step instructions) with food images, built to train models that match photos to recipes. A subset carries nutritional information. Access goes through a request form and is limited to universities and research institutions.
Access
- Cost
- By applicationFree, but download links are emailed after a request form; access is granted only to universities and research institutions.
- License
- Open, non-commercial onlyResearch use only
- Sign-up
- Application required
- How you get it
- Bulk downloadJSON, LMDB, JPEG
What's in it
- One row is
- One record per recipe, with linked images
- Covers
- Mostly English-language recipe websites
- Years
- Released 2017; extended 2019
- Updated
- One-time release
- Size
- About 1 million recipes; image archive in ten tar files of roughly 210 GB each
Quality
The standard benchmark for recipe and food-image research, but it is a static scrape with noisy ingredient text and murky upstream rights.
- Scraped content from third-party recipe sites
- Ingredient quantities are unstructured text
- Not updated since 2019
Connects to
- USDA FoodData Central: Map ingredients to nutrient profiles to estimate a recipe's nutrition
- NYPL What's on the Menu?: Compare home recipes with historical restaurant menus
Questions it could answer
- Which ingredients most often appear together, and how does that differ by cuisine?
- How has the typical number of steps in a recipe changed across source sites?