PLANS: per-image outputs, image lists, reference values and analysis scripts (version 2.0)

Data and scripts accompanying: Nakagawa S, Yamamoto A. Development and Evaluation of an On-Device Dietary Self-Monitoring System Using Open-Weight Vision–Language Models: Model Selection and Evaluation on Three Public Datasets. Submitted to JMIR mHealth and uHealth (2026). PLANS = Personal-device Local Assessment of Nutrients System. The iOS/iPadOS application (PLANS-iOS) and the research batch application are archived separately (GitHub: https://github.com/shnkgw-rincom/PLANS-iOS). What changed from version 1.0. Version 1.0 of this record accompanied an earlier manuscript that used a non-public image set (NutriImage) together with SNAPMe. That manuscript was withdrawn by the authors because the conditions of use of NutriImage could not be confirmed. Version 2.0 contains no NutriImage material. All evaluations were re-run or re-aggregated on three public image sets with reference values: SNAPMe, ACETADA and Nutrition5k. Files of version 1.0 that concerned NutriImage are not carried over; the SNAPMe files and the acceptance test are carried over unchanged. Contents: image lists; reference values derived from the three releases; per-image model outputs (126 files: 15 open-weight models on a server in BF16, FP8 checkpoints, declared-weight conditions, 4-bit runs on a desktop computer and a tablet, a same-stack BF16 control, and six commercial models); declared-weight inputs; the acceptance test of PLANS-iOS; aggregated tables; and the run and analysis scripts. See README.md and MANIFEST.tsv in the archive. Images are not redistributed; obtain them from the original releases. Outputs and derived tables are released under CC BY-NC 4.0, scripts under the MIT License.