ReciFine: Large Silver-standard Dataset

ReciFine is a large, silver-standard dataset derived from the RecipeNLG corpus and enriched with fine-grained semantic annotations.

Scale

• 2.2+ million recipes
• 97+ million entity mentions
• Token-level annotations across all recipe instructions

Entity Distribution

Entity Type	Frequency	Top Entities
Food (F)	30,199,222	salt, water, sugar, butter, flour
Tool (T)	10,384,889	bowl, oven, pan, saucepan
Duration (D)	3,700,982	10 min, 5 min, 30 min
Quantity (Q)	4,476,287	remaining, all, 2, half, 1
Chef Action (Ac)	30,854,542	add, bake, mix, stir, cook
Discont. Ac (Ac2)	2,199,530	together, to taste, to a boil
Food Action (Af)	3,023,358	cool, stand, set, combined
Tool Action (At)	705,504	comes out, stand, set
Food State (Sf)	7,331,595	hot, tender, smooth, browned
Tool State (St)	5,022,448	large, medium, small, 350°

Table 1 Frequencies of entities for each ReciFine entity type, along with their most frequent entities.

ReciFine Annotation Schema

The ReciFine follow the The Case Study Paper annotation scheme and include 10 fine-grained entity types annotated at token level within recipe instructions.

Tag	Name	Definition
F	Food	Edible items; includes both raw ingredients and intermediate products
T	Tool	Cooking tools such as knives, bowls and pans
D	Duration	Time durations used in cooking (e.g., 20 minutes)
Q	Quantity	Quantities associated with ingredients
Ac	Action by chef	Verbs denoting deliberate actions by the cook (e.g., bring in “Bring the mixture to a boil.”)
Ac2	Discontinuous Ac	Non-contiguous parts of compound chef actions (e.g., to a boil in “Bring the mixture to a boil.”)
Af	Action by food	Verbs where food is the agent (e.g., melt, boil)
At	Action by tool	Verbs pertaining to a tool's action (e.g., grind, beat)
Sf	Food state	Descriptions of food's physical state (e.g., chopped, soft)
St	Tool state	Descriptions of tool state or readiness (e.g., preheated, greased, covered)

Table 2: Entity types in the English Recipe Flow Graph corpus.

Evaluating Automatic Annotations

To assess the reliability and validity of the ReciFine silver-standard annotations, we compared model-generated labels from RecipeBERT and RecipeRoBERTa against human annotations on 500 randomly selected recipes. The models achieved strong agreement with human labels, confirming that the large-scale automatically annotated ReciFine corpus closely aligns with expert annotation quality. This finding mirrors the results observed on the ERFG corpus, further supporting the consistency and robustness of ReciFine’s finely annotated extractions.