RecipeDB2
A large-scale structured recipe dataset constructed by the CoSyLab laboratory at IIIT-Delhi. It contains 128,942 recipes and 35,474 ingredients spanning 32 regions across 99 countries, integrating ingredient entity recognition, nutritional linking, and cultural annotations to make recipe data computationally accessible.
Timeline (last 90 days)
RecipeDB2 is a structured recipe dataset built by a research team at IIIT-Delhi that decomposes each recipe into typed ingredient entities, anchors them to a reference nutrition database, and annotates geocultural and dietary context information.
RecipeDB2 compiles 128,942 recipes covering 35,474 ingredients from 32 regions and 99 countries.
RecipeDB2 uses a Transformer-based named entity recognition model to parse ingredient phrases, splitting them into seven culinary attributes.
RecipeDB2 links ingredients to the U.S. Department of Agriculture (USDA) reference nutrition table via a BERT embedding strategy.
On a manually adjudicated set of the 200 most frequent ingredients, the nutrition linking F1 score reaches 87.90.
RecipeDB2 provides 148 nutritional parameters for each mapped ingredient.
RecipeDB2 uses a random forest classifier to extend 34 ingredient categories to the entire ingredient vocabulary.
RecipeDB2 assigns a dietary style (e.g., vegetarian, vegan) to each recipe using a set of deterministic, conservative rules.
The RecipeDB2 dataset has been publicly released through IIIT-Delhi's CoSyLab laboratory at cosylab.iiitd.edu.in/recipedb2.
The evaluation set for F1 = 87.90 is limited to the 200 most frequent ingredients, and the research team has not provided independent evaluation figures for linking quality on long-tail, low-frequency ingredients.
1 more timeline events
All Facts (11)
RecipeDB2 adopts a layered engineering approach: NER handles linguistic parsing, BERT embeddings handle entity linking, random forest handles category propagation, and a rule set handles dietary style determination, rather than using a single large model for everything.
50%UnverifiedRecipeDB2 compiles 128,942 recipes covering 35,474 ingredients from 32 regions and 99 countries.
50%UnverifiedRecipeDB2 uses a Transformer-based named entity recognition model to parse ingredient phrases, splitting them into seven culinary attributes.
50%UnverifiedRecipeDB2 links ingredients to the U.S. Department of Agriculture (USDA) reference nutrition table via a BERT embedding strategy.
50%UnverifiedOn a manually adjudicated set of the 200 most frequent ingredients, the nutrition linking F1 score reaches 87.90.
50%UnverifiedRecipeDB2 provides 148 nutritional parameters for each mapped ingredient.
50%UnverifiedRecipeDB2 uses a random forest classifier to extend 34 ingredient categories to the entire ingredient vocabulary.
50%UnverifiedRecipeDB2 assigns a dietary style (e.g., vegetarian, vegan) to each recipe using a set of deterministic, conservative rules.
50%UnverifiedRecipeDB2 is a structured recipe dataset built by a research team at IIIT-Delhi that decomposes each recipe into typed ingredient entities, anchors them to a reference nutrition database, and annotates geocultural and dietary context information.
50%UnverifiedThe evaluation set for F1 = 87.90 is limited to the 200 most frequent ingredients, and the research team has not provided independent evaluation figures for linking quality on long-tail, low-frequency ingredients.
50%UnverifiedThe RecipeDB2 dataset has been publicly released through IIIT-Delhi's CoSyLab laboratory at cosylab.iiitd.edu.in/recipedb2.
50%