ingredient-embeddings multilingual-ai recipe-generation food-technology

All of human cooking compressed into 2 megabytes

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

arxiv.org

May 27, 2026

2 min read

🔥🔥🔥🔥🔥

66/100

Summary

Epicure consists of three sibling skip-gram ingredient embeddings retrained on a multilingual recipe corpus. The dataset includes 4.14 million recipes sourced from 11 different platforms and covers seven languages.

Key Takeaways

Epicure is a family of three skip-gram ingredient embeddings retrained from scratch on a multilingual recipe corpus of 4.14 million recipes from 11 sources in seven languages.
The embeddings normalize raw ingredient strings to 1,790 canonical entries using an LLM-augmented pipeline.
Epicure includes a 203,508-edge ingredient-ingredient NPMI graph and an 80,019-edge typed FlavorDB ingredient-compound graph.
Three Metapath2Vec variants are developed, differing in their random-walk schemas: Cooc, Chem, and Core, each positioned on a chemistry-vs-recipe-context spectrum.

Read original article

View story →

Coverage & discussions

Hacker News · arxiv.org

All of human cooking compressed into 2 megabytes ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Mixed

Positives

The idea of compressing human ingredients into a simplified format could revolutionize how we understand flavor combinations, making it easier for chefs to innovate.
The concept of using a limited set of ingredients to explore culinary possibilities is intriguing and could serve as a valuable resource for future cooking applications.

Concerns

The claim of encompassing 'all of human cooking' is misleading, as the dataset heavily favors English and Chinese sources, neglecting significant global cuisines.
The lack of representation from major culinary traditions like Italian and Mexican raises concerns about the dataset's comprehensiveness and applicability.