idoyaaran/mise-recipes

🍳 Mise Recipes

A synthetic dataset of 10,000 cooking recipes across 8 cuisines, generated with a pretrained language model and cleaned/validated with a deterministic EDA pipeline. Built for Mise, a "Duolingo-for-cooking" app: given a cuisine, it recommends similar recipes and generates a progressive cooking lesson.

Intro to Data Science final project, Reichman University.

Dataset at a glance

• 10,000 recipes, 1,250 per cuisine (perfectly balanced).
• 8 cuisines — regional: italian, mexican, mediterranean, indian, levantine; style: comfort_food, farm_to_table, vegan.
• Modality: text. Language: English.
• Avg 5.7 ingredients and 4.8 steps per recipe.

How it was generated

• Model: Qwen/Qwen2.5-3B-Instruct (fp16), run on a GPU via batched generation.
• Method: one recipe per prompt, varying cuisine × dish type × difficulty across the run (not a single mega-prompt). The model was instructed to return strict JSON matching the schema below, flag core/pantry ingredients, and keep dietary tags consistent with the ingredients.
• The generation notebook is included in this repo.

Schema

One recipe per row:

field	type	notes
id	string	{A/B}_{cuisine}_{000001}
title	string	dish name
cuisine	string	one of the 8 cuisines
dish_type	string	breakfast, main, side, dessert, …
difficulty	string	beginner / intermediate / advanced (also the curriculum level)
ingredients	list	{name, quantity, is_basic_local}
steps	list[str]	ordered instructions
techniques	list[str]	e.g. sauteing, roasting, baking
dietary_tags	list[str]	vegan, vegetarian, gluten_free, dairy_free, nut_free, contains_pork, contains_shellfish, spicy
time_minutes	int	total time
servings	int

EDA & data-quality notes

Synthetic data contains model mistakes; we identified and corrected them during EDA (see the EDA notebook in this repo).

• Dietary-tag correction: the model's self-reported dietary_tags matched the actual ingredient list only ~9% of the time (it both contradicted — e.g. dairy_free on a recipe with parmesan — and omitted applicable tags). We recompute all dietary tags deterministically from the ingredients, so the published tags are 100% consistent.
• Diversity limitation: the generator over-favors iconic dishes. There are 0 exact-duplicate recipes, but many share a title with different content (e.g. 129 distinct "falafel balls", 118 "tiramisu"). Unique titles per cuisine range from ~575 (Italian, least diverse) to ~975 (farm-to-table). These are kept as legitimate variations.
• Difficulty mix: ~23% beginner-weighted toward beginner/intermediate by design.

Load it

import pandas as pd
df = pd.read_parquet("hf://datasets/<your-username>/mise-recipes/recipes_clean.parquet")

Files in this repo

• recipes_clean.parquet — the cleaned dataset (use this).
• 02_merge_and_eda.ipynb — merge + EDA + dietary-tag corrector.
• generation notebook — the synthetic-data-generation code.