Dataset
Restaurant Pilot Meal Set (N=32, pilot)
Summary
A pilot dataset of 32 restaurant-served meals with reference kcal values derived either from published chain-restaurant nutrition disclosures (verified by dismantle-and-weigh) or from dietitian-led recipe reconstruction. Released as a companion to the restaurant-extension preprint (DAI-PRE-2026-02); not yet a full reference set.
Description
The Restaurant Pilot Meal Set is a small dataset of 32 restaurant-served meals collected in one metropolitan area during February 2026. It is the pilot-stage companion to the Initiative’s restaurant-extension protocol (preprint DAI-PRE-2026-02) and is released to support methodological critique prior to the full N=200 extension planned for 2026-Q4.
Unlike the Weighed-Meal Reference Set v1.0 (mini-180), where per-meal kcal is a deterministic function of weighed ingredients and FDC entries, restaurant meals carry irreducible reference-value uncertainty. Each meal in this pilot is released with an explicit uncertainty half-width rather than as a point-value ground truth. Users are expected to propagate this uncertainty in any downstream analysis.
Composition:
| Source path | N meals | Description |
|---|---|---|
| Path A (published-nutrition, verified) | 18 | Chain restaurants that publish per-menu-item nutrition; verified by post-hoc dismantle-and-weigh. |
| Path B (recipe reconstruction) | 14 | Independent restaurants without published nutrition; reconstructed by a research dietitian. |
| Total | 32 |
Schema
| Field name | Type | Description |
|---|---|---|
meal_id | string | Unique identifier, format RPM-<3-digit> |
restaurant_type | enum | ”chain” / “independent” |
restaurant_id_anonymised | string | Opaque identifier (e.g., “CHAIN-03”) |
menu_item_name | string | As printed on the menu |
capture_date | date (ISO) | Date the meal was served |
reference_path | enum | ”path_a_published” / “path_b_reconstruction” |
reference_kcal | float | Best-estimate reference kcal for the served portion |
uncertainty_halfwidth_pct | float | Plus/minus uncertainty as a percentage of reference_kcal |
served_weight_g | float | Dismantle-verified served weight in grams |
published_portion_weight_g | float (nullable) | Chain-published portion weight, where applicable |
portion_scaling_applied | boolean | True if reference_kcal was scaled from published by weight ratio (Path A only) |
reconstruction_notes | string (nullable) | Path B: summary of assumptions made during recipe reconstruction |
ingredients_observed | list (nullable) | Path B: itemised observed ingredients with weighed grams where available |
CSV and JSON formats provided.
Provenance / collection methodology
Path A. For chain restaurants that publish per-menu-item kcal and portion weight, the menu item was ordered, served, photographed (photos not released publicly), then dismantled and re-weighed in a nearby prep area under controlled conditions. If the dismantled served weight deviated from the published portion weight by more than 10%, the published kcal was scaled by the weight ratio (portion_scaling_applied = true). Uncertainty half-width for Path A is typically +/- 4.0%, driven by residual composition uncertainty within the scaled portion and by the precision of the prep-area balance.
Path B. For independent restaurants that do not publish nutrition information, the meal was dismantled and weighed in the prep area; a research dietitian reconstructed the recipe from the menu description, the dismantled weights, and visible preparation cues. Ingredients were mapped to USDA FDC Foundation Foods entries (April 2026 snapshot). A conservative uncertainty half-width was assigned per reconstruction, based on the number and nature of assumptions required — ranging from +/- 7.2% (straightforward salads and grilled items) to +/- 18.9% (deep-fried items where the cooking-fat component could not be directly observed).
Restaurant identities are anonymised. The sampling was purposive (not random) — the pilot aimed for coverage of a range of restaurant types rather than statistical representativeness.
Known limitations
- Small N and single metropolitan area. Thirty-two meals from one region cannot support population-level inference. This is a pilot.
- Purposive sampling. Restaurants were chosen for cooperation and accessibility, not random selection.
- Higher per-meal uncertainty than metabolic-kitchen references. The uncertainty half-widths in this pilot are substantially larger than those of the Weighed-Meal Reference Set v1.0, and must be propagated in any downstream application-accuracy analysis.
- Chain-disclosure drift. Published chain-nutrition figures may change over time; the figures used here are those available on 2026-02-15 through 2026-02-28.
- No imagery published. Photographs were captured but are not released with this dataset in order to minimise restaurant identifiability.
Versioning
This is the pilot release (N=32) dated March 2026. It is intended as a companion to the restaurant-extension preprint. A full N=200 extension is planned for 2026-Q4 and will be released as a separately versioned dataset (Restaurant Reference Set v1.0), not as an update-in-place of this pilot.
How to access
CSV and JSON files are available for direct open download from the Initiative’s datasets page. No access procedure is required beyond the CC BY 4.0 attribution requirement. Photographic material is not released.
How to cite
Rivera S, Weiss H. (2026). Restaurant Pilot Meal Set (N=32, pilot). The Dietary Assessment Initiative.
Users citing this pilot should make explicit that it is a pilot and should report the per-meal uncertainty half-width alongside any derived accuracy figure.
License
Creative Commons Attribution 4.0 International (CC BY 4.0).
Cite this dataset
Rivera S, Weiss H. (2026). Restaurant Pilot Meal Set (N=32, pilot). The Dietary Assessment Initiative.
Keywords
restaurant meals; pilot dataset; reference kcal; dismantle-and-weigh; recipe reconstruction; external validity; dietary assessment