rahul7star/food-recipes

Food.com Multimodal Recipe Dataset (15K)

Dataset Summary

Property	Value
Total samples	~15,000
Modalities per sample	2 — PNG recipe card image + Markdown text
Image format	PNG, 300 DPI, A4 aspect ratio
Source dataset	Food.com Recipes and User Interactions (Kaggle)
Raw recipe pool	~231,637 recipes
License	See source dataset license

---

Intended Use Cases

This dataset was designed to support the following downstream research and engineering tasks:

• Multimodal RAG — Visual and text-based recipe retrieval using vision-language embedding models (e.g., nvidia/llama-nemotron-embed-vl-1b-v2)
• VLM fine-tuning — Structured document understanding, ingredient extraction, and nutritional reasoning
• Cross-modal retrieval — Ingredient-based, nutrition-aware, and image-based recipe search
• Document layout understanding — Recipe cards follow a consistent yet visually diverse template across 10 colour schemes and 4 chart styles

---

Data Sources

Two raw files from the Food.com Recipes and User Interactions Kaggle dataset were used as input:

File	Description	Raw Size
RAW_recipes.csv	Recipe metadata, ingredients, steps, nutrition, tags	~231,637 rows
RAW_interactions.csv	User ratings and free-text reviews per recipe	Several hundred thousand rows

---

Dataset Construction Pipeline

The dataset was built using a six-stage reproducible pipeline.

Stage 1 — Data Loading and Exploratory Analysis

Both raw CSVs were loaded via pandas. A quality audit was performed covering missing value counts, duplicate recipe detection, and distribution analysis of n_steps and n_ingredients. Coverage was measured across a range of filtering thresholds (5, 8, 10, 12, 15, 18, 20, 25) to inform downstream cutoff selection. All rows containing null values were dropped prior to further processing.

Stage 2 — Interaction Signal Aggregation

The interactions table was aggregated at the recipe level. For each recipe, the mean star rating across all reviewers was computed, individual review texts were collected into a list (nulls excluded), and the total number of distinct user ratings was counted. This aggregated table was left-joined onto the recipes DataFrame on recipe_id. The original id column was subsequently dropped, with recipe_id serving as the canonical identifier for all downstream stages.

Stage 3 — Filtering and Sampling

The full merged corpus of ~231,637 recipes was reduced to the final ~15,000 subset using the following criteria applied in sequence:

Step	Criterion	Rationale
1	Sort by rating descending	Prioritises higher-quality, well-reviewed recipes
2	n_ingredients ≤ 15	Ensures ingredient lists render cleanly on a single card
3	n_steps ≤ 15	Prevents step overflow across card pages
4	num_ratings > 10	Requires a minimum review signal to filter low-confidence recipes
5	Drop remaining null rows	Final cleanliness pass before checkpoint serialisation

Two columns unused downstream (contributor_id, submitted) were dropped at this stage. The filtered subset was saved as an intermediate CSV checkpoint.

Stage 4 — Review Text Cleaning

Raw review strings from the interactions table contained HTML entities, encoding artifacts, and short noise strings. Each string was normalised in the following order: HTML entity decoding, non-ASCII character stripping, whitespace and newline collapsing, and filtering of strings shorter than 5 characters. Cleaned reviews were stored back into the DataFrame as Python lists. Up to 5 reviews were used for the Markdown text modality; up to 10 (sampled deterministically by row index) were used for the recipe card image.

Stage 5 — Markdown Text Modality Generation

Each row was serialised into a structured Markdown string constituting the text modality of the sample. The schema covers the following fields in a fixed order:

• Recipe title, recipe ID, cook time
• Star rating and total review count
• Description paragraph
• Comma-separated ingredient list
• Numbered preparation steps
• Compact nutrition summary: Calories, Total Fat, Sugar, Sodium, Protein, Saturated Fat (Note: the 7th nutrition element, Total Carbs, is intentionally omitted from the compact summary)
• Top 5 cleaned user reviews

This Markdown string is stored in the markdown column and is intended to be embedded directly alongside the image for multimodal retrieval.

Stage 6 — Recipe Card Image Modality Generation

Each recipe was rendered as a styled single-page A4 PDF using a custom multi-module Python pipeline (config.py, utils.py, charts.py, builder.py, pipeline.py). The PDF was subsequently rasterised to PNG at 300 DPI using PyMuPDF (no Poppler dependency required). Images are stored at data/output/images/recipe_id_<ID>.png.

Visual variation system: To maximise visual diversity across the 15,000 cards, each recipe card is assigned a unique visual style determined by its row index. The system combines 10 colour schemes with 4 nutrition chart types, yielding 40 base style combinations. Assignments are seeded by row index, ensuring full reproducibility across pipeline runs.

The 10 colour schemes are: WarmSpice, FreshHerb, OceanBreeze, MidnightChef, BerryFusion, SunsetGrill, SlateCuisine, TropicalBowl, RusticBakery, and CrimsonFeast.

The 4 nutrition chart variants are: vertical bar chart, horizontal bar chart, radar chart, and table-only (no chart).

Card layout: Each card contains the following sections — header (title, cook time, star rating), ingredient list, numbered preparation steps, nutrition panel (with the assigned chart variant), and a review section (up to 10 cleaned reviews). If a required field is missing or unparseable, that section is omitted rather than rendered as a placeholder.

Multi-page overflow handling: If a card exceeded one page, the pipeline automatically reduced the review count and, if necessary, applied progressive font scaling until the card fit within a single page or a minimum scale threshold was reached. Only the first page was exported to PNG.

Failed renders: 12 recipe IDs failed during image generation and were excluded from the final dataset. Additional rows where the expected PNG was not found on disk were also excluded, producing the final ~15,000 sample count.

---

Dataset Schema

Column	Type	Description
name	string	Recipe name (title-cased)
recipe_id	int64	Unique recipe identifier
minutes	int64	Total preparation and cook time in minutes
description	string	Recipe description text
tags	list[string]	Descriptive tags (up to 8 per card)
steps	list[string]	Ordered preparation steps (numbered, sentence-cased)
n_steps	int64	Number of preparation steps (≤ 15)
ingredients	list[string]	List of ingredient strings
n_ingredients	int64	Number of ingredients (≤ 15)
nutrition	struct	Nutritional values as a labelled struct of float32 fields
nutrition.calories	float32	Energy in kilocalories
nutrition.total_fat_pdv	float32	Total fat as % Daily Value
nutrition.sugar_pdv	float32	Sugar as % Daily Value
nutrition.sodium_pdv	float32	Sodium as % Daily Value
nutrition.protein_pdv	float32	Protein as % Daily Value
nutrition.saturated_fat_pdv	float32	Saturated fat as % Daily Value
nutrition.carbs_pdv	float32	Total carbohydrates as % Daily Value
rating	float32	Mean star rating across all reviewers (1.0–5.0)
num_ratings	float32	Total number of user ratings (> 10 for all samples)
markdown	string	Full structured Markdown recipe (text modality)
image	Image	Rendered recipe card PNG at 300 DPI (image modality)

---

Nutrition Values — Important Note

Nutritional values are sourced directly from the original Food.com dataset and are expressed as Percent Daily Values (PDV) rather than absolute gram weights, with the exception of Calories (reported in kilocalories). PDV values represent the percentage of the recommended daily intake contributed by one serving of the recipe, based on a 2,000-calorie reference diet.

---

Data Quality Summary

Metric	Value
Raw recipe pool	~231,637
After quality filtering (n_steps ≤ 15, n_ingredients ≤ 15, num_ratings > 10)	~15,700
Excluded — failed image renders	12
Excluded — image file missing on disk	Varies
Final dataset size	~15,000 samples
Modalities per sample	2 (PNG image + Markdown text)
Minimum ratings per recipe	> 10
Image resolution	300 DPI (A4)

---

Recommended Embedding Model

This dataset was designed for use with nvidia/llama-nemotron-embed-vl-1b-v2, a 1B-parameter vision-language embedding model producing 2048-dimensional vectors from both image and text inputs. Recipe card images are intended to be embedded directly as document images, while Markdown strings can be embedded as text or incorporated into a combined multimodal input.

---

Repository Structure

data/
├── all csv files/
│   ├── RAW_recipes.csv              # Original Kaggle recipes file
│   ├── RAW_interactions.csv         # Original Kaggle interactions file
│   └── recipes_15k_samples.csv      # Intermediate filtered CSV checkpoint
└── output/
    └── images/
        └── recipe_id_<ID>.png       # Rendered recipe card images (300 DPI PNG)

---

Dependencies

Library	Purpose
pandas	Data loading, merging, filtering, and serialisation
ast	Safe parsing of stringified Python lists from CSV columns
re, html	Review text cleaning and HTML entity decoding
reportlab	PDF layout generation for recipe cards
PyMuPDF (fitz)	PDF-to-PNG rasterisation at 300 DPI (no Poppler required)
Pillow (PIL)	Image loading, colour conversion, and JPEG encoding
numpy	Numerical type handling and NaN detection
datasets	Hugging Face Dataset construction and Hub upload
tqdm	Progress tracking across the generation pipeline

---

Citation

If you use this dataset in your research, please cite both the original Food.com source dataset and this curated multimodal version.

Original Source — Food.com Recipes and User Interactions

@inproceedings{majumder2019generating,
  title     = {Generating Personalized Recipes from Historical User Preferences},
  author    = {Majumder, Bodhisattwa Prasad and Li, Shuyang and Ni, Jianmo and McAuley, Julian},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing
               and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
  year      = {2019}
}

This Dataset

@misc{ghosh2025food15k,
  author       = {Tipto Ghosh},
  title        = {Food.com Multimodal Recipe Dataset (15K)},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/tiptoghosh/food-recipes-15k}},
  note         = {Curated multimodal recipe dataset with rendered recipe card images and
                  structured Markdown text representations.
                  Project repository: \url{https://github.com/Tipto-Ghosh/OmniChef-Nexus}.
                  Author profile: \url{https://scholar.google.com/citations?user=tiptoghosh}}
}