FoodClock Labs is the model-building side of FoodClock, a calorie-budget app that is built on Open Food Facts. We publish open models and datasets for one job: reading what is printed on food packaging, from phone photos, in any language.

What is here

• Etiquette-1-8B: a vision-language model (Qwen3-VL-8B-Instruct, fine-tuned and merged, Apache 2.0) that takes the photos of a product and returns one JSON record of what the pack says: names per language, brand, net quantity, the nutrition table row by row as printed, the ingredient list per language, allergen, certification, origin and recycling marks, and a tag per photo. It reads; it does not estimate, derive or complete. The exact prompt and a short script are on the card.
• etiquette-printed-truth: the exam the model had to pass. 100 products, what a person recorded as printed on each pack from the photos ("not printed" is a value), the model's replies, and a scorer that is one standard-library Python file. No GPU, no model, no network needed to re-run the number.

How we work

• Locked exams, published whole. A model is judged on sets of products drawn once and never trained on. The public hundred is here with the replies and the scorer; a second hundred stays private and is scored in totals only, so the public set cannot quietly become the training set.
• Reading, not guessing. A blank where nothing is printed beats a plausible number. Inventions (a value returned where the pack prints none) are counted and reported on their own.
• Everything traceable. Every card names the Open Food Facts photos it was read from. Photos stay on Open Food Facts under their own licence (CC BY-SA 3.0); the data is ODbL. We credit the contributors who photographed and typed in the packs, because none of this exists without them.

More about the model, in plain words: foodclock.com/app/labs/etiquette. Questions, packs it reads wrong, things you built with it: hello@foodclock.com.