Palate

Open data

250 canonical dishes, measured.

When we shared our look at when ingredients enter a recipe, the most common question was: can we have the recipes? Yes. This is a hand-picked sample of our corpus — 250 dishes we'd call canonical, six or seven per cuisine across all 36 cuisines in our hand-authored curated set. Paella, pho bo, doro wat, khachapuri, mole poblano, cacio e pepe.

Every dish carries the engine's core measurement: an 18-note flavor profile, the salt-fat-acid-heat balance, per-100g macros, timing, and the gram-normalized ingredient list. It's the same flavor data behind every hexagon on this site — in a form you can actually load into pandas.

Licensed CC BY-NC 4.0 — analyze it, visualize it, write about it, teach with it; just credit Palate with a link back to this page, and get in touch first for commercial use. Every row also links to its live palate_url so readers of whatever you build can see the dish's X-ray themselves.

What's inside

The columns, in plain language

GroupColumnsWhat it tells you
Identityrecipe_id · title · cuisine · category · region · palate_urlWhich dish this is, where it's from, and a link to its live flavor X-ray.
Timing & efforttotal / active / passive_minutes · effortHow long it takes and how much of that is hands-on.
Dietary & macrosvegetarian · vegan · contains_* · calories…fiber (per 100 g)Best-effort flags and per-100g macros — useful for filtering, not allergen-safe.
Flavor (18 notes)flavor_umami … flavor_marine_briny · flavor_potencyThe engine's objective flavor profile — the same 18 notes behind every hexagon on the site.
Balancebalance_salt · fat · acid · thermal_heatThe salt-fat-acid-heat backbone, normalized 0–1.
Ingredientsingredients · ingredient_linesEngine-normalized ingredient masses in grams, plus the hand-written recipe lines.

The spread

36 cuisines, evenly weighted

Six or seven dishes from each — so no single food culture dominates the sample, and cross-cuisine comparisons start from level ground.

  • American Southern
  • Argentine
  • Brazilian
  • British
  • Cajun/Creole
  • Chinese
  • Cuban
  • Egyptian
  • Ethiopian
  • Filipino
  • French
  • Georgian
  • German
  • Greek
  • Hungarian
  • Indian
  • Indonesian
  • Italian
  • Jamaican
  • Japanese
  • Korean
  • Lebanese
  • Malaysian
  • Mexican
  • Moroccan
  • Nigerian
  • Persian
  • Peruvian
  • Polish
  • Portuguese
  • Russian
  • Scandinavian
  • Spanish
  • Thai
  • Turkish
  • Vietnamese

How it's measured

Measurements, not opinions

The flavor numbers are deterministic: language models do perception and structure only — reading a recipe into ingredients and steps — and never emit a flavor number. From there, the engine propagates each ingredient's flavor through the cooking steps (mass-aware, technique by technique) the same way every time. Run it twice, get the same profile twice.

These 250 recipes come from our hand-authored curated set, so the ingredient lines are ours to share. Dietary flags and macros are engine-derived estimates — good for filtering and analysis, not allergen decisions.

The flavor layer is open on purpose. The engine measures more than this — texture structure, and process features like Maillard development and flavor melding, derived from how a dish is actually cooked — and that layer stays inside the product. If your project needs it, ask us.

See the data journalism →Browse the full corpus →See it mapped in the Atlas →

Build something with it.

Want more — the full corpus, the ingredient flavor lexicon, pairing affinities, or a slice cut for a specific question? Tell us what you're building through the feedback button anywhere in the app, or find us on Reddit where this started. We read everything.

Explore the corpus →