Graph-Based Dataset for Cooking Activity Recognition from Recipe Structures

This dataset supports research on the procedural understanding of cooking recipes and fine-grained culinary activity modeling. It is derived from the Recipe1M corpus and was generated through an automated pipeline combining large language model annotation, controlled culinary verb normalization, and procedural variant generation. The repository contains three complementary datasets: recipes_instructions_annotated.csv : Annotated recipe instructions where each instruction is mapped to standardized culinary actions using a controlled vocabulary of cooking verbs, with labels indicating whether the action corresponds to a physical gesture. recipes_sequences_variants.csv : Structured action sequences representing multiple procedural variants of recipes, capturing alternative execution paths and ingredient-informed preparation steps. recipe_graphs.db : A database containing graph representations of recipes, where nodes correspond to normalized culinary actions and edges represent procedural transitions between actions. Together, these resources enable research on procedural modeling, action sequence analysis, and graph-based representations of cooking processes for applications such as activity recognition and intelligent kitchen systems.