PLANS: per-image outputs, inputs and analysis scripts (version 1.0)
Abstract
Per-image model outputs, input files and analysis scripts accompanying: Nakagawa S, Yamamoto A. Development and Evaluation of an Offline Vision-Language Model System for Estimating Nutrients from Meal Photographs on Personal Devices: The PLANS (Personal-device Local Assessment of Nutrients System) Study. Submitted to Journal of Medical Systems (2026). PLANS is a system in which a small open vision-language model (Qwen3-VL, 4-bit) runs entirely on a personal Apple device and estimates energy, protein, fat, carbohydrate and salt equivalent from a meal photograph without network access. This record contains: (1) per-image outputs of the tablet runs (iPad Air, NutriImage 887 and SNAPMe 1,134 images, base and persona conditions), the desktop true-weight runs (2B/4B/8B), the eleven declared-weight-error runs (4B) and the Gemini 3.8 Flash true-weight run; (2) the declared-weight input files; (3) the reference tables; (4) the acceptance-test records of the PLANS-iOS application against the research batch application on the same iPhone, with 23 of the 26 test photographs; (5) the values of Tables S3-S6 and of the 68 local cells; and (6) the aggregation, figure and run scripts. The desktop base-condition runs are in the related record https://doi.org/10.5281/zenodo.22917651. The source code of PLANS-iOS v0.1.0-paper2 is archived separately: [DOI of PLANS-iOS].Outputs, tables, reference tables and photographs: CC BY 4.0. Scripts: MIT License. See README.md for the file layout, the aggregation rules and how to run the scripts.
