MenuVision Dataset

A synthetic dataset of 10,037 restaurant dishes spanning 20 world cuisines, built as the knowledge base for MenuVision — an AI dish-recommendation system that fetches live restaurant menus and matches them, by semantic similarity, against these dishes to infer spice level, flavor profile, cuisine, and a clean description for each real menu item.

The dataset was generated with large language models and then cleaned, validated, and enriched through an iterative EDA cycle. This card documents that EDA in full: what the raw generation produced, the data-quality problems it revealed, and the fixes that turned it into the final dataset.

---

Dataset at a glance

Property	Value
Rows (dishes)	10,037
Core columns	9
Primary cuisines	20
Null values (final)	0
Duplicate dish names	0
Split	train

Schema

Column	Type	Description
dish_id	int	Unique identifier
dish_name	string	Dish name (all unique)
cuisine	string	One of 20 primary cuisines (fusion/regional labels remapped to their base cuisine)
description	string	Short natural-language description (~13 words)
ingredients	string	Comma-separated ingredient list (~5 items)
dietary_tags	string	Pipe-separated multi-label tags (e.g. vegan|gluten-free|nut-free)
spice_level	string	mild | medium | hot | very hot
flavor_profile	string	e.g. savory, sweet, rich, spicy
price_range	string	budget | mid | premium

The description and ingredients fields carry the most free-text signal, but all eight non-ID columns are concatenated into a single composite string for embedding (see the EDA "Text quality" section).

---

How the dataset was generated

Dish records were produced by prompting an instruction-tuned LLM to emit structured JSON, one batch at a time, across all 20 cuisines. Choosing the generator and stabilizing the prompt was itself an EDA-driven process.

Model bake-off. Several open models were evaluated before settling on the generator:

Model	Outcome
TinyLlama-1.1B	97% duplicate dish names — rejected
SmolLM2-360M	Weak instruction-following — rejected
Phi-3-mini	Impractically slow generation — rejected
Llama-3.1-8B	Coherent, low duplication — selected (via HF Inference API)
Qwen2.5-7B	Used for the later top-up batch

This bake-off is the empirical basis for generating with a served Llama-3.1-8B rather than a small self-hostable model — the small models that fit modest hardware simply produced unusable output.

Prompt iteration. The generation prompt went through four documented rounds:

Iteration	Change	Problem it solved
1	Batch size 50	40% of batches truncated the JSON
2	Batch size reduced to 25	Duplication was climbing toward 30%
3	Theme rotation (12 themes) + variation nudges	Residual duplication
4	repair_truncated_json fallback	Recovered 35–54 dishes per failed batch

By the final iteration the duplicate rate had dropped from ~30% to 1–2%, and generation was stable.

---

Exploratory Data Analysis

The EDA was run as an iterative loop: analyze → find a defect → fix generation or clean the data → re-validate. The sections below report what was found and what was done about it.

1. Data quality — nulls

The raw dataset had a very low null rate (< 0.2% overall): 17 missing ingredients (0.18%), 5 missing dietary_tags (0.05%), and 3 missing spice_level (0.03%). These came from the generator occasionally returning an empty string for a field rather than omitting it. They were filled deterministically (ingredients → unknown, dietary_tags → none, spice_level → medium), bringing nulls to 0.

2. Data quality — schema normalization

Validating each categorical column against its allowed vocabulary exposed the biggest quality issue in the raw data — the model frequently used out-of-vocabulary values:

Column	Invalid values	Share of dataset
price_range	4,933	52.4%
flavor_profile	3,600	38.2%
spice_level	433	4.6%
cuisine	173	1.8%

For example, price_range came back as low/high instead of budget/premium, and spice_level appeared as moderate, mellow, not hot, etc. These were normalized to the canonical vocabularies (the 173 non-standard cuisine values were remapped to their primary cuisine — see Section 4). Crucially, the original values were preserved in *_raw columns so that nothing informative was lost before embedding. After normalization, invalid spice_level, price_range, and cuisine counts were all 0.

3. Dietary hallucination detection and repair

The most consequential finding: the model's dietary_tags were internally inconsistent with its own ingredients. Cross-checking tags against ingredients flagged a ~12% hallucination rate — dishes labeled vegan that listed meat (457) or dairy (329), and vegetarian dishes listing meat (535), for ~1,149 conflicts in total.

Because incorrect dietary tags are a safety issue in a recommender, the tags were re-derived from scratch from the ingredient lists (and dish names) rather than trusted as generated. This also moved the data from mostly single-label to genuinely multi-label:

Tag	Share after re-derivation
nut-free	85.6%
halal	82.0%
kosher	79.7%
dairy-free	64.9%
gluten-free	59.6%
vegetarian	42.3%
vegan	18.4%
low-carb	12.4%
high-protein	9.5%

After the fix, 98.4% of dishes carry two or more tags. A residual audit still flagged ~52 vegan-labeled dishes containing subtle animal products (e.g. foie gras, venison), which are noted as edge cases. This "re-derive from real signals rather than trust generated labels" principle carries directly into the live app, where dietary tags are re-inferred from real menu descriptions.

4. Distributions

Cuisine — the generator initially produced 89 distinct cuisine values, including (approxemately) 40 fusion and regional labels (e.g. Israeli-Japanese, Peruvian-Chinese, Sichuan, French Late Night). These 173 rows (1.8%) were remapped to their primary cuisine — fusion labels to their first-named base (Israeli-Japanese → Israeli), Chinese regional styles to Chinese, and so on. The rationale: the primary cuisine carries the strongest cultural signal for the embedding model, and 5–7-row micro-categories aren't usable by the recommender. After remapping, the dataset has exactly 20 balanced cuisines (coefficient of variation 0.18), from Mexican (655) down to Moroccan (359).

Before remapping — 89 cuisines, including the fusion/regional sprawl:

After remapping — the clean 20 primary cuisines:

Spice level (final): mild 40.2%, medium 25.6%, hot 17.7%, very hot 16.5%.

Flavor profile: savory 27.7%, sweet 11.9%, rich 7.8%, spicy 6.7%, with a long tail of creamy, umami, sour, smoky, and others.

Price range: skewed toward mid (~45–48%), with budget and premium making up the rest.

Dietary tags: none was the most common single value before multi-label re-derivation (~24%); afterward the distribution is the multi-label one in Section 3.

5. Text quality (embedding readiness)

Each dish is embedded as a composite text string built from eight fields — dish_name, description, ingredients, cuisine, dietary_tags, spice_level, flavor_profile, and price_range — concatenated into one embedding_text (e.g. "Pad Ka Pow. Thai basil stir-fry with chicken and chilies. Ingredients: … Cuisine: Thai. Dietary: … Spice: hot. Flavor: savory. Price: mid."). So the embedding captures a dish's full profile, not just its description. The two free-text fields carry the most signal, so their quality was checked closely:

• Descriptions: mean 13.3 words (std 4.0); 96.7% fall in the ideal 8–50 word range; 3.2% too short, 0.1% too long. This tight, consistent length is well suited to sentence embeddings.
• Ingredients: mean 5.2 items (std 0.7); 98.5% within the 4–6 target range — strong instruction adherence. Only 17 rows (0.18%) had unknown ingredients.

6. Cross-field insights

A spice × flavor breakdown showed the data is internally coherent rather than random:

• Savory dominates at every spice level (38–51%).
• Sweet falls as spice rises (mild 29% → very hot 12%) — desserts cluster at mild, as expected.
• Spicy flavor rises with spice level (mild 6% → very hot 22%) — the model correctly tied the spicy flavor to higher heat.

Price by cuisine matched intuition too: French is the most premium-heavy cuisine (42.7% premium dishes) and Japanese the most budget-heavy (36.5% budget).

7. Uniqueness

After the generation fixes, the dataset has 0 duplicate dish names — every one of the ~9,400 core names (pre-top-up) was unique, a direct result of the batch-size and theme-rotation prompt changes.

---

Known limitations

• Mild-spice skew. mild is over-represented (~40% vs. a 25% uniform ideal). This is left unbalanced: it plausibly reflects real menus (which skew mild/medium), and hot/very hot still have ~1,600+ dishes each — enough for good recommendations.
• Cuisine granularity. Fusion and regional cuisines were collapsed into the 20 primary cuisines by design, so some granularity (e.g. distinct Chinese regional styles) is intentionally lost in exchange for balanced, RecSys-usable categories.
• Synthetic origin. All content is model-generated. The EDA above corrects the systematic errors it introduced (dietary hallucinations, schema drift), but the dishes are not scraped from real menus.
• Dish-name realism. Some synthetic dish names are implausible or non-existent — the generator preserved real culinary roots (e.g. koshari, musakhan, misir wat, qabuli) but sometimes invented specifics or over-qualified them. This is inherent to LLM generation. It does not affect the live app, which recommends from real restaurant menus and uses this dataset only as a semantic knowledge base, never as a source of dish names shown to users.

---

How the dataset is used in MenuVision

At runtime the app embeds each dish's composite profile text with paraphrase-multilingual-mpnet-base-v2 (selected in Part 3). When a live Wolt menu is fetched, each real dish is matched by cosine similarity to its nearest dish here, and the matched record supplies spice level, flavor profile, cuisine, and a clean description. Dietary tags are re-inferred from the real menu text, following the same principle established during EDA. The dataset therefore functions as a semantic knowledge base, not just a static table.

Files in this repository

• data/ — the dataset (Parquet)
• dish_embeddings.parquet — precomputed 768-dim embeddings (Part 3)
• embedding_metadata.json — winning embedding model + metadata
• plots/ — EDA figures referenced in this card
• Part 1 (data generation) and Part 2 (EDA) notebooks