Palate

A measured toy

Guess the dish

Set something on the counter and the question of what you’re making narrows. Flour tortilla, cumin, lime, and it is basically settled. Salt and water, and you have said nothing yet. Add ingredients below and watch the guess move.

The model behind it is the simplest thing that works: it counts which ingredients turn up in which kinds of dish across 3,018 labelled recipes, then reads your board back through those counts. It never sees a quantity, an order, or a step — only what is on the counter.

In one paragraph

A Naive Bayes over ingredient presence names the form of a dish (22 forms, from soup and stir-fry to taco/wrap) correctly 45.9% of the time, with the right answer in its top three 73.1% of the time. Measured with 5-fold cross-validation over 3,018 recipes using a vocabulary of the 400 most common ingredients; the majority-class baseline is 13.5%. The model reads presence only: never quantity, order, or technique. A board whose best guess does not clear 35% is published as “off our map” rather than as an answer.

What’s on the counter

It thinks you're making

taco/wrap, at 99%. The runner-up is curry at 0%.

  1. taco/wrap99%
  2. curry0%
  3. hearty main0%
  4. stew0%
  5. main (other)0%

How sure are we

98% decided

Decided. On boards like this the model is rarely surprised. This is the spread of the guess, not its correctness — a confident model can still be wrong.

What each one decided

Some ingredients settle the question by themselves; most leave it wide open. This is the entropy of P(dish form | that one ingredient), on a 0–1 scale where 1 means fully decisive — measured over the same 3,018 recipes. Vanilla extract decides it. Salt promises nothing.

  • flour tortilla0.79

    Leans taco/wrap — 86% of its 35 recipes, 24.2× the base rate.

  • lime0.30

    Leans taco/wrap — 28% of its 159 recipes, 8.0× the base rate.

  • cumin0.22

    Leans taco/wrap — 19% of its 395 recipes, 5.3× the base rate.

  • cilantro0.21

    Leans taco/wrap — 16% of its 273 recipes, 4.5× the base rate.

  • ground beef0.19

    No form is over-represented enough to name.

How this was measured

Every recipe in Palate’s corpus carries a dish-form label: 22 of them, from soup and stir-fry to taco/wrap and an honest main (other). A multinomial Naive Bayes was fitted over ingredient presence across the 400 most common ingredients, then scored with 5-fold cross-validation: 45.9% top-1 and 73.1% top-3, against a majority-class baseline of 13.5%. That is a real signal from arithmetic you could do on paper — and it is also the floor a better model has to clear.

The tables ship with the page (multinomial naive bayes over ingredient presence (pairing_key vocabulary, alpha=0.5, top-400 by df)), so the guess is computed in your browser. Nothing you type is sent anywhere, stored, or logged.

What it can’t do

  • It reads only what you bought. A soup and a stew are the same chicken and the same carrot; what separates them is how much liquid stayed in the pot. Nothing in an ingredient list records that, so those are the pairs it confuses.
  • The labels were written once, by a language model, and frozen. A careful reading against a fixed rubric, not a ground truth — and a human taxonomist would disagree at exactly the boundaries where the errors sit.
  • The vocabulary is 400 ingredients. Anything outside it is ignored, never approximated. That is why an unusual board goes quiet instead of confident.
  • The corpus is Western-skewed. Commitment measures convention, not necessity: rice noodles look decisive here because this corpus mostly puts them in one kind of dish. Against a corpus written in Hanoi they would decide far less.

The full write-up (the commitment ladder, the confusable pairs, and how few ingredients it takes before the picture resolves) is in the Instrument Room: Which ingredients decide the dish.

Now the harder question: would you like it?

The same corpus, measured on eighteen notes of flavor instead of on form. Twenty-five questions, and you see your result before anyone asks you to sign up.

Take the taste test →

Cite this

Palate, “Guess the dish,” https://palate.kitchen/guess — include the date you accessed it.

The measurements on this page are free to quote with attribution to Palate. The downloadable 250-dish sample dataset on /data is licensed CC BY-NC 4.0.