FoodClockLabs/etiquette-printed-truth

Etiquette printed truth

100 food products, and what is printed on their packs, recorded by a person from the photos. This is the exam that gated the release of Etiquette 1 (8B), FoodClock's label-reading model, together with the model's replies on it and the scorer that turns the two into the numbers on the model card. Anyone can re-run the score, and anyone can score another model on the same cards.

Most label datasets score a model against a catalogue record. A catalogue record is a curator's line: the name is what someone typed, salt and sodium are a pair with one computed from the other, a value can come from a different pack than the photo shows. This set records what the pack prints, from the photos themselves, with "not printed" as a value of its own, so a model that invents a number where the pack has none is caught, and a model that stays silent there is right.

What is in it

file	lines	what
cards.jsonl	100	one product per line: its Open Food Facts code, the photos the person read it from (ids and URLs), and the printed truth per field
replies/run13.jsonl	100	Etiquette 1 (8B)'s reply for each card, raw text and the parsed JSON object, on exactly those photos
score.py		the scorer: python3 score.py cards.jsonl replies/run13.jsonl prints the model-card numbers

The photos are not in this repository. They belong to the Open Food Facts contributors who took them (ODbL for the data, CC BY-SA 3.0 for the images); each card gives their URLs on images.openfoodfacts.org, and the ids are stable. Fetch them politely (a named user agent, one at a time) if you want to run your own model.

A card

{
 "code": "0028400079143",
 "country": "United States",
 "photos": [
  {
   "iid": 1,
   "url": "https://images.openfoodfacts.org/images/products/002/840/007/9143/1.jpg"
  },
  {
   "iid": 4,
   "url": "https://images.openfoodfacts.org/images/products/002/840/007/9143/4.jpg"
  }
 ],
 "truth": {
  "product_name": {
   "value": "Chocolate Brownie",
   "printed": true
  },
  "brand": {
   "value": "Grandma's",
   "printed": true
  },
  "quantity": {
   "value": "70.8 g (2 1/2 oz)",
   "printed": true
  },
  "basis": {
   "value": "serving: 1 Cookie (35g)",
   "printed": true
  },
  "energy_kcal_100g": {
   "value": "160",
   "printed": true
  },
  "salt_100g": {
   "value": null,
   "printed": false
  },
  "sodium_mg": {
   "value": "115.0",
   "printed": true
  },
  "front_names": {
   "value": {
    "en": "Chocolate Brownie"
   },
   "printed": true,
   "source": "human"
  },
  "generic_names": {
   "value": null,
   "printed": false,
   "source": "human"
  }
 }
}

Fields: product_name is the name line as the front presents it; brand as printed; quantity the net quantity as printed; basis the basis of the first value column of the nutrition table (100g, 100ml, serving: <size as printed>, or 1 L (...) for a mineral water); energy_kcal_100g, salt_100g (g) and sodium_mg (mg) the figures in that first column, each only if that row is printed - never converted from one another; front_names and generic_names the name on the front and the descriptive or legal denomination, one entry per language or script printed (ISO 639-1 keys, und for a coined name). "printed": false means the field is not printed, or not legible, in the ticked photos. Every value is verbatim from the ticked photos; a word cut by the photo edge is left out.

source on the two name fields: human where the person typed or confirmed the entry, claude where the entry was pre-filled by a Claude model reading the photos and the person has not (yet) confirmed it - the model card says the same. Every other field is human-confirmed on every card.

A reply

{
 "code": "0028400079143",
 "run": 13,
 "model": "FoodClockLabs/Etiquette-1-8B",
 "photos": [
  1,
  4
 ],
 "reply": "{\n  \"kind\": \"food\",\n  \"product_name\": {\n    \"en\": \"Chocolate Brownie\"\n  },\n  \"generic_name\": {\n    \"en\": \"COOKIES\"\n  },\n  \"brand\": \"Grandma's\",\n  \"quantity\": \"2...",
 "reading": {
  "kind": "food",
  "nutriments": {
   "energy_kcal_100g": 160,
   "fat_100g": 6.0,
   "saturated_fat_100g": 2.5,
   "carbohydrates_100g": 23.0,
   "sugars_100g": 12.0,
   "fiber_100g": 1.0,
   "proteins_100g": 2.0
  },
  "product_name": "Chocolate Brownie",
  "brand": "Grandma's",
  "product_names": {
   "en": "Chocolate Brownie"
  },
  "generic_names": {
   "en": "COOKIES"
  },
  "photos": [
   [
    "front"
   ],
   [
    "nutrition"
   ]
  ],
  "quantity": "2 1/2 OZ. (70.8 g)",
  "ingredients": {},
  "looped": [],
  "basis": {
   "per": "serving",
   "header": "Per serving",
   "size": "1 Cookie (35g)"
  },
  "minerals_per_l": {},
  "sodium_mg": 115.0,
  "allergens_printed": null,
  "may_contain_printed": null,
  "marks": [],
  "marks_printed": [],
  "packaging": {},
  "truncated": false,
  "rows": {
   "energy_kcal": {
    "value": 160,
    "unit": "kcal"
   },
   "fat": {
    "value": 6,
    "unit": "g"
   },
   "saturated_fat": {
    "value": 2.5,
    "unit": "g"
   },
   "carbohydrates": {
    "value": 23,
    "unit": "g"
   },
   "sugars": {
    "value": 12,
    "unit": "g"
   },
   "fiber": {
    "value": 1,
    "unit": "g",
    "qualifier": "<"
   },
   "proteins": {
    "value": 2,
    "unit": "g"
   },
   "sodium": {
    "value": 115,
    "unit": "mg"
   }
  },
  "qualifiers": {
   "fiber_100g": "<"
  }
 }
}

reply is the model's text as returned (greedy decoding, 2,048 new tokens, the prompt and system message published with the model). reading is that text parsed into one object: the JSON object between the first { and the last }, with two repairs for a reply the token limit cut - a string that repeats a phrase three times in a row is closed at its last separator, and an unterminated string is closed rather than the whole list dropped. The scorer reads reading; reply is there so you can parse it your own way.

The numbers

python3 score.py cards.jsonl replies/run13.jsonl - no GPU, no model, no network; the two files and the standard library. It prints, for run 13:

printed-truth score: 100 cards, 981 fields | agree 814 differ 40 missing 45 invented 42 | 83.0%

and a per-field table. How a field is judged: printed and read equal = agree; printed and read other = differ; printed and nothing read = missing; not printed and nothing read = agree (silence is the right answer); not printed and a value read = invented. Numbers agree within 2% (or 0.05). Text agrees after normalisation (accents, case, punctuation) when equal, one inside the other, or sharing 60% of the shorter side's words; a net quantity agrees after unit and e-mark normalisation ("220 g" = "220g e"). The per-language name fields are judged per entry; a right text filed under another slot (brand / front / generic) or another language key is a "slot" filing disagreement - the headline percentage counts it as not agreeing, the per-field percentages leave it out. The script's header says the same, in full.

Where the cards come from

The 100 products are a fixed draw from the Open Food Facts products FoodClock itself had read and pushed to the database by late September 2026 - a holdout no training run ever saw (every product in this set, and in our other exams, is excluded from the training export by code). The mix follows where those products came from: about half are Moroccan packs (Arabic and French on the same pack), a quarter French, the rest from Belgium, India, Algeria, the UK, the US, Argentina and elsewhere, with ingredient lists in 12 languages. Countries are from the catalogue record.

The truth was recorded on 29 September - 2 October 2026 on a page that shows the product's photos and, per field, three answers: as the catalogue has it, edited (typed from the pack), or not printed. The person ticks the photos the card was read from; the model is sent only those photos. Fields pre-filled by a Claude model from the photos were confirmed or corrected by the person; the text fields were then audited against the ticked photos (32 changes on 100 cards, mostly descriptors or transliterations the photos do not show, now removed). The numbers on the model card use this audited truth.

Using it to score another model

1. For each card, fetch its photos (in that order) and run your model with the prompt of your choice - the Etiquette prompt and system message are in the model repository and give the field names score.py looks for.
2. Parse the reply to one JSON object and write {"code": ..., "reading": <object>} per line.
3. python3 score.py cards.jsonl yours.jsonl --examples.

The scorer looks at reading.product_names, generic_names, product_name, brand, quantity, basis.per, nutriments.energy_kcal_100g, nutriments.salt_100g and sodium_mg; any other key is ignored. A fair comparison sends the same photos the person ticked, and nothing else.

Licence and attribution

component	holder	licence
cards.jsonl (the printed-truth records), README.md	FoodClock Labs 2026	CC BY-SA 4.0 - see LICENSE
replies/run13.jsonl (the model's replies)	FoodClock Labs 2026	CC BY-SA 4.0
score.py	FoodClock Labs 2026	Apache-2.0
The products, their codes and the photos the cards were read from (linked, not included)	Open Food Facts and its contributors	database ODbL 1.0, contents DbCL 1.0, photos CC BY-SA 3.0

Contains information from Open Food Facts, which is made available here under the Open Database License (ODbL). Every card names the Open Food Facts photos it was read from by their URLs; the photos themselves are not in this repository and remain the work of the contributors who took them, under CC BY-SA 3.0. The cards are transcriptions of text printed on packs, recorded by a person; they are released under CC BY-SA 4.0, the compatible successor of the photos' licence, so share-alike is met on any reading. Open Food Facts is a trademark of the Open Food Facts association; FoodClock Labs is not affiliated with or endorsed by it.

What this asks of you: credit FoodClock Labs for the cards and Open Food Facts and its contributors for the products and photos; share derived card sets under CC BY-SA 4.0. Nothing here restricts commercial use.

Nothing here is dietary or medical advice; the cards record what a pack prints, which is not always what it contains.