Tameka/themealdb_clean

Dataset Card for TheMealDB Recipes (Cleaned)

Dataset Summary

This dataset contains 789 recipes collected from the free TheMealDB API. Each recipe has:

• its name, category (e.g. Dessert, Beef, Vegetarian) and country of origin
• cooking instructions
• a list of ingredients with their measures
• tags
• links to an image, a YouTube video and the original recipe website.

Two fields were added during this project: has_video and num_ingredients. The dataset is intended for educational purposes. It was built for Assessment and is used for text classification with a Transformers pipeline and semantic recipe search with FAISS.

Supported Tasks

• sentence-similarity / feature-extraction (semantic search). Recipe names, ingredients and instructions can be turned into text embeddings and indexed with FAISS, so users can find similar recipes from a plain-English query (e.g. "spicy vegetarian dinner"). Retrieval quality can be checked by whether the top-k results are relevant to the query.
• text-classification. Used in Task 2 with a zero-shot pipeline (facebook/bart-large-mnli) to classify recipes as "vegetarian" or "made with meat or fish" from the name and instructions, checked against a rule-based label from the ingredients (82.9% accuracy on the test split). strCategory (14 classes) could also be used as a label, but the classes overlap and are very unbalanced (see Discussion of Biases).

Languages

The text is in English (BCP-47: en), mostly British English: "coriander", "plain flour", metric measures. Some dish names are in their original language (e.g. "Hete bliksem", "Chakchouka").

Dataset Structure

Data Instances

A typical record (instructions shortened):

{
  "idMeal": "53159",
  "strMeal": "Chorizo, potato & cheese omelette",
  "strCategory": "Pork",
  "strCountry": "Spain",
  "strInstructions": "step 1\r\nCook the potato in boiling water for 8-10 mins or until tender. Drain and allow to steam-dry. Heat oil in an omelette pan, add chorizo and cook for 2...",
  "strMealThumb": "https://www.themealdb.com/images/media/meals/0y6uvc1763258983.jpg",
  "strTags": "Cheesy",
  "strYoutube": "https://www.youtube.com/watch?v=K2Ba5BWYXUU",
  "strSource": "https://www.bbcgoodfood.com/recipes/chorizo-potato-cheese-omelette",
  "dateModified": "2025-11-16 02:07:20",
  "has_video": true,
  "num_ingredients": 6,
  "ingredients": [
    {"ingredient": "potatoes", "measure": "1 small"},
    {"ingredient": "olive oil", "measure": "1 teaspoon"},
    {"ingredient": "chorizo", "measure": "50g"},
    {"ingredient": "egg", "measure": "3"},
    {"ingredient": "parsley", "measure": "Chopped"},
    {"ingredient": "cheddar cheese", "measure": "25g"}
  ]
}

Each record is one recipe, and records are independent of each other. There are no explicit links between recipes.

Data Fields

Field	Type	Description	Missing
idMeal	string	TheMealDB's unique meal ID. Maps directly to https://www.themealdb.com/meal/{idMeal}	0
strMeal	string	Meal name	0
strCategory	string	One of 14 categories. Label for text classification	0
strCountry	string	Country the dish is associated with (63 countries)	0
strInstructions	string	Cooking method. Input text for the pipeline and embeddings	0
strMealThumb	string	Link to the meal image on TheMealDB (images are not stored here)	0
strTags	string	Comma-separated tags, e.g. "Spicy,Bun,Calorific"	591
strYoutube	string	YouTube video link	197 (stored as "")
strSource	string	Original recipe website	36 (11 null, 25 "")
dateModified	timestamp	When the record was last changed on TheMealDB	302
has_video	bool	Added. True if strYoutube is not empty	0
num_ingredients	int	Added. Number of ingredient lines (2–20, median 10)	0
ingredients	list of {ingredient: string, measure: string}	Cleaned. Each ingredient (lower-cased) with its measure, in recipe order. Input text for embeddings	0

Data Splits

	Train	Test
Recipes	631	158

The split is 80/20 and random, made with train_test_split(test_size=0.2, seed=42). The fixed seed makes it reproducible. Both splits come from the same source and were cleaned the same way. Across the whole dataset, instructions average 975 characters (median 842), and recipes average 10.3 ingredients.

Dataset Creation

Curation Rationale

The dataset was built to practise a full data workflow for an AI project: collect → augment → clean → publish → analyse → search. Recipes were chosen for four reasons:

• They suit the later tasks. They combine short labels (category, country) with free text (instructions, ingredients). That suits both text classification and semantic search.
• The source is open and documented. TheMealDB has a public, documented API with a free test key, so the collection is easy to reproduce.
• The privacy risk is low. Recipes describe food, not people.
• The size is right. At around 800 records the dataset fits a free Kaggle notebook, yet it is large enough to show meaningful patterns.

Source Data

Initial Data Collection and Normalization

Collection

• The data comes from TheMealDB API v1, search-by-first-letter endpoint: https://www.themealdb.com/api/json/v1/1/search.php?f={letter}, using the public test key 1.
• A Python requests loop sends one request for each letter a–z, so the search covers every meal whose name starts with any letter. No keywords or other filters were used.
• Letters with no meals return {"meals": null}. These are skipped so the loop doesn't crash.
• The loop waits 0.5 seconds between requests to avoid overloading the free service. The whole collection takes about 15 seconds.
• Records were saved to meals.jsonl and loaded with datasets.load_dataset("json", ...). This gave 790 records with 54 fields.
• Date collected: 2026-10-12

Augmentation: two fields were added using .map().

Field	How	Why
has_video	lambda: True if strYoutube is not empty	Shows how many recipes have a video tutorial (592 of 789)
num_ingredients	Function counting the non-empty strIngredient1–20 slots	A simple measure of recipe complexity

Inspection before cleaning: 790 rows and 56 columns. The problems found were:

• 40 sparse ingredient/measure columns
• 3 columns that were empty for every row
• 190 missing strArea values
• one junk record
• the same ingredient spelt with different capitals (1,144 spellings for 946 real ingredients).

Cleaning: every change, what was done and why.

Step	Method	What was done	Why	Result
1	map()	Combined strIngredient1–20 and strMeasure1–20 into one ingredients column: a list of {"ingredient", "measure"} dictionaries. Ingredient names were stripped of spaces and lower-cased	Most of the 40 columns were blank. One column is easier to read and use for semantic search. Lower-casing merges "Garlic" and "garlic". Fixed dictionary keys let Hugging Face store the column as a typed structure	List length equals num_ingredients for every recipe
1a	decision	Kept duplicate ingredients	They come from separate parts of a recipe (e.g. garlic in both the fritters and the filling). Keeping them stays faithful to the source and keeps each measure paired with its ingredient	—
2	remove_columns()	Removed the 40 original ingredient/measure columns	Their information is now in ingredients	57 → 17 columns
3	remove_columns()	Dropped strMealAlternate, strImageSource, strCreativeCommonsConfirmed	Empty for all 790 rows	17 → 14 columns
4	remove_columns()	Dropped strArea	Missing for 190 rows and used mixed naming ("British" vs "France"). strCountry holds the same information with 0 missing	14 → 13 columns
5	decision	Kept strTags even though 591 rows are empty	198 meals have useful tags (e.g. "Spicy", "Keto") for semantic search. Most meals were simply never tagged	—
6	filter()	Kept only recipes with instructions of at least 50 characters	I inspected every recipe under 150 characters. Only "Bread omelette" ("Make and enjoy") had no usable method. The next shortest (about 77 characters) were real recipes, so 50 was a safe cut-off	790 → 789 rows
7	train_test_split()	80/20 split, seed=42	Gives held-out data for later evaluation; reproducible	631 / 158

Principles followed:

• Drop a column only if it is empty or duplicated elsewhere.
• Keep partly filled columns that are useful, and document the gaps.
• Never invent values. dateModified, strYoutube and strSource were left with gaps for this reason.

Tools: Python, requests, pandas, Hugging Face datasets, Kaggle notebooks.

Who are the source language producers?

The recipes were written by people, not generated by a machine. TheMealDB's content is added by its community of users. Many recipes are adapted from published recipe websites, linked in strSource. The largest single source is BBC Good Food (277 recipes, 35%). Nothing is known about the contributors' identities or demographics, and they are not recorded in the data.

Annotations

Annotation process

No manual annotation was done.

• Labels from the source. strCategory, strCountry and strTags were assigned by TheMealDB contributors.
• Fields added in code. has_video and num_ingredients were computed by code from existing fields. No human judgement was involved.

Who are the annotators?

Category, country and tag labels were assigned by TheMealDB's community contributors. Their selection criteria and identities are unknown. The derived fields were created by the dataset curator's code.

Personal and Sensitive Information

• No personal information. The dataset describes recipes, not people. It contains no names, email addresses, user IDs, contributor details, locations of individuals, or other personal information. No individual can be identified, directly or indirectly. The Privacy Act 1988 (Cth) and the Australian Privacy Principles therefore have little application, and no anonymisation was needed.
• Countries are linked to dishes, not people. strCountry connects dishes to national cuisines, not to people's origins.
• Public links only. All links point to public websites (TheMealDB, YouTube, recipe sites).

Considerations for Using the Data

Social Impact of Dataset

Positive. Recipe search tools can help people:

• find meals from what they have at home, which reduces food waste
• discover dishes from other cultures
• find recipes that match how they like to eat (e.g. vegetarian).

Risks

• Health. Recipes have no nutrition or allergen information and have not been checked by experts. A tool built on this data could recommend unsuitable recipes to someone with allergies or medical dietary needs. Do not use this data for food-safety or dietary advice.
• Cultural representation. Search and classification will favour the cuisines that appear most often, so less-represented food cultures will be harder to find.

Discussion of Biases

• Uneven coverage of countries. The dataset covers 63 countries, but they are not evenly represented:
  • The top 10 countries hold 42% of recipes (333 of 789), and the United Kingdom has the most (59).
  • Ten countries have fewer than 5 recipes (e.g. Bahamas 1, Angola 2).
  • Much of Africa, the Middle East and South Asia is thinly covered.
• One dominant source. 35% of recipes come from BBC Good Food, so the style, ingredients and measures are mostly British.
• Unbalanced categories. Dessert has 167 recipes, while Vegan has 7 and Goat has 2.
• Simplified labels. Each dish gets one country, though many dishes are shared across a region.

Steps taken: these biases were measured and documented. They were not corrected, because rebalancing the data would mean removing or inventing records. Proposed: report results by country and category, use F1 score per class for classification, and collect more recipes from under-represented regions.

Other Known Limitations

• Crowd-sourced and not verified. Quantities, cooking times and labels may contain errors.
• Data was added in batches. All 190 records with no strArea come from 28 countries, nearly all starting with A–D (e.g. Afghanistan, Brazil, Cambodia). This suggests a newer batch of records added with a different format.
• Inconsistent missing values and leftover headings. strYoutube stores missing values as "", strSource uses both null and "", and 199 recipes have "step 1"-style headings in strInstructions (copied from the source website). These are stored as-is in this dataset, but were corrected during preprocessing in Notebook 3 before embedding (see Task 3, Section 2).
• Tags are one string. strTags is a comma-separated string, not a list.
• Small dataset. 789 rows is suitable for learning and demonstration, not for production models.
• Snapshot. TheMealDB changes over time, so collecting again may give different data.

Additional Information

Dataset Curators

Tameka, student at North Metropolitan TAFE, for Assessment. There was no funding.

Licensing Information

TheMealDB doesn't use a standard open licence such as Creative Commons. The licence is therefore set to other and links to the TheMealDB Terms of Use. The key terms are:

• "You can scrape, copy and modify any content returned from the API, as long as you use the official end points." This dataset was collected only through the official API.
• You "must mention us as the source of the data" and "should link back to our website where appropriate."
• The free key is for "testing and personal projects".
• "You cannot publish apps to an appstore unless you are a paid subscriber", and "You cannot resell our API in any way without specific permission."

Attribution: Data from TheMealDB.

Intended use: education and non-commercial research only.

Copyright

• Many recipe instructions are adapted from other websites. Those sites may own copyright in the original text under the Copyright Act 1968 (Cth). This dataset is used for study and research (fair dealing), and strSource credits the original author.
• No images are stored. strMealThumb only links to TheMealDB. strCreativeCommonsConfirmed was empty for every record, so image licences could not be confirmed. Images should not be reused without checking first.

Storage, Security and Versioning

Credentials. The Hugging Face token is stored in Kaggle Secrets (UserSecretsClient().get_secret("HF")). It never appears in the notebook code or outputs.

Where things are stored:

Item	Classification	Location	Access
Raw data	Public	Hub repo Tameka/themealdb_raw (Parquet, 790 rows)	Public
Cleaned data	Public	Hub repo Tameka/themealdb_clean (Parquet, train 631 / test 158)	Public
Notebook 1: collection and cleaning	Internal	Private Kaggle notebook	Author and assessors only
Notebook 2: Transformers pipeline	Internal	Private Kaggle notebook	Author and assessors only
Notebook 3: FAISS semantic search	Internal	Private Kaggle notebook	Author and assessors only
Outputs (charts, results tables)	Internal	Saved inside each notebook, with the code that produced them	Author and assessors only
Documentation	Public	This README.md	Public

Why the notebooks are not public: they are assessment work. Under the Data Access and Classification Policy (§3, §5 least privilege / need-to-know), they are classified as Internal and shared only with assessors, to protect academic integrity. The dataset itself is classified Public because it holds no personal information and comes from an open API.

Versioning.

• Datasets: each Hugging Face repo is a Git repository. Every upload or edit is saved as a commit, visible under "Files and versions". The raw and cleaned data are kept in separate repos, so the original is never overwritten.
• Notebooks: each Kaggle "Save Version" creates a numbered version with its outputs, so earlier runs can be recovered. The final versions were submitted.

Organisational Policy Alignment

This project follows the Koorliny Kaatijin policy set (Version 1.0, effective 2025-10-15).

Policy	Relevant requirement	How this project complies
Data Access and Classification Policy	§3: classify all information assets (Public / Internal / Confidential / Restricted). §4: label datasets clearly	Classified as Public. The source is a public API, the data holds no personal information, and disclosure causes no harm. This card is the label. Because it is Public, release on the public Hugging Face Hub is appropriate
Data Access and Classification Policy	§9: assess third-party or cloud services before use	Kaggle and the Hugging Face Hub were the platforms directed for this course. Only Public data is uploaded to them
Privacy Policy	§4–5: collect personal information only when necessary, and use it only for its purpose. §6: rules for cross-border disclosure	No personal or sensitive information is collected (§3 definitions), so the APPs and cross-border rules are not triggered. The Hugging Face servers are overseas, which is acceptable because the data is Public
Privacy Policy	§7: keep data accurate and complete	Data quality steps are documented: integrity checks, junk-record filter, known issues listed
Information Security and Acceptable Use Policy	§3: accounts are individual, and sharing credentials is prohibited. §3: store data according to its classification	The Hugging Face token is personal, kept in Kaggle Secrets, and never written into code, outputs or this card
Information Security and Acceptable Use Policy	§4: do not bypass security controls	Only the official, documented API endpoints were used, with a 0.5-second delay between requests. There was no scraping around access controls
AI Use Policy	§7: confirm lawful basis and licensing for all datasets; respect copyright and terms of use	TheMealDB Terms of Use were reviewed and are quoted under Licensing Information. Attribution is given, use is non-commercial, strSource credits original authors, and no images are stored
AI Use Policy	§5 and §7: record data sources for reproducibility; maintain data lineage	Provenance, every cleaning step and both dataset versions (raw → clean) are documented in this card
AI Use Policy	§5: conduct bias evaluations. §8: treat AI outputs as drafts until reviewed	Biases in coverage and categories are measured under Discussion of Biases. Pipeline and search outputs were reviewed: the Task 2 classifier was checked against a rule-based answer key (82.9% accuracy, with 13 dangerous errors analysed), and Task 3 search results were scored by hand (precision@5), with coverage gaps confirmed in code
Data Retention and Disposal Policy	§4: project documentation (non-regulated) is kept for 5 years after project closure	The dataset, notebooks and card are kept for the assessment period, then retained or deleted according to this schedule
Data Retention and Disposal Policy	§6: cloud data is deleted using the provider's certified procedures. §7: keep disposal records	Disposal means deleting the repositories through Hugging Face's repository settings, deleting the Kaggle notebooks, and recording the dataset name, date and method
Data Breach Response Procedure	§3: a breach includes compromised data integrity. §5: contain the breach and revoke compromised credentials	No personal data is held, so the risk of a notifiable breach (NDB scheme) is negligible. If the token were exposed: revoke it on Hugging Face immediately, issue a new one, check the commit history for unauthorised changes, and report it within 24 hours (InfoSec Policy §3)

Risks and Mitigations

Risk	Action taken	Future mitigation
Biased results from uneven coverage	Measured and documented	Report results by group; collect from under-represented regions
Errors in crowd-sourced recipes	Junk record removed; warning against dietary use	Spot-check a sample against strSource
Copyright in recipe text and images	strSource attribution kept; images not stored; non-commercial use	Remove a recipe if a rights holder asks
Breaching TheMealDB terms	Official endpoints only; attribution given; 0.5-second delay between requests	Get a paid key if use becomes commercial
Token leak	Token kept in Kaggle Secrets and never printed	Follow the Data Breach Response Procedure: revoke and regenerate the token, check commit history, and report within 24 hours
Keeping data longer than needed	Retention period set (Data Retention and Disposal Policy §4)	Delete the repos and notebook after the retention period, and record how they were disposed of
Source data changes	Raw snapshot kept in themealdb_raw	Record the date of each collection
Users rely on AI to judge if a recipe is vegetarian (Task 2)	13 of 76 meat dishes (17%) were wrongly called vegetarian; documented with a warning	Never use for dietary, religious or allergy decisions without a human check; flag low-confidence predictions (< 0.65)
Search returns results for cuisines that are not in the data (Task 3)	Found: 0 Nigerian and 0 First Nations recipes, yet 5 results were still returned	Warn users when no recipe matches the requested country; collect recipes from under-represented regions with cultural consultation
Search ignores "vegetarian" and "without" (Task 3)	Documented with examples	Add rule-based filters after the search

Citation Information

@misc{themealdb_clean_2026,
  author       = {Tameka Carmichael},
  title        = {TheMealDB Recipes (Cleaned)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Tameka/themealdb_clean}},
  note         = {Data sourced from TheMealDB, https://www.themealdb.com}
}

References

• TheMealDB API guide: https://www.themealdb.com/docs_api_guide.php
• TheMealDB Terms of Use: https://www.themealdb.com/terms_of_use.php
• Hugging Face LLM Course, "Creating your own dataset": https://huggingface.co/learn/llm-course/chapter5/5
• Hugging Face, "Dataset cards": https://huggingface.co/docs/hub/datasets-cards
• Gebru et al. (2021), Datasheets for Datasets: https://arxiv.org/abs/1803.09010
• Privacy Act 1988 (Cth): https://www.legislation.gov.au/C2004A03712/latest
• Copyright Act 1968 (Cth): https://www.legislation.gov.au/C1968A00063/latest

Contributions

Thanks to TheMealDB and its contributors for the source data.