FoodAudit-AG

FoodAudit-AG is an auditable workflow for food-additive compliance-state assessment under GB 2760-2024.

Partial dataset release

This repository publishes part of the benchmark dataset: 150 of 450 recipes.

本仓库公布部分数据集：完整基准共 450 条配方，本次公开其中 150 条。

The public release contains 150 recipe-level records and 403 additive-level judgments with labels, category paths, and regulatory evidence fields. The remaining 300 records are part of the internally curated benchmark and have not been cleared for unrestricted redistribution.

Public-slice selection

The released slice preserves one third of every predefined difficulty stratum in the full benchmark: L1, 30 of 90; L2, 45 of 135; L3, 60 of 180; and L4, 15 of 45. The slice is intended for transparency and mechanism-level verification and is not claimed to be a random or commercially representative sample or a complete substitute for the full benchmark.

See data/partial_dataset/v1/README.md and DATA_CARD.md.

Repository structure

• data/partial_dataset/v1/: public partial dataset in CSV and JSONL formats.
• data/partial_dataset/v1/hard_noise_probe_30.csv: 30-case hard-noise diagnostic subset derived from the public partial dataset.
• results/robustness/: aggregate hard-noise robustness results.
• resources/kg_snapshot/v1/: frozen category tree, permission rules, and category synonyms.
• src/: FoodAudit-AG backend snapshot used as the revision baseline.
• app/: review-oriented Streamlit prototype.
• scripts/: graph import, candidate generation, evaluation, statistics, and release builders.
• docs/category_anchoring.md: exact candidate-generation, ranking, threshold, output-resolution, and failure-handling specification.
• docs/robustness_hard_noise.md: hard-noise perturbation construction and summary results.
• CODE_FREEZE.json: checksums for the revision-baseline code and resources.

Release status

• Public dataset: partial-dataset-v1.0.0.
• Full benchmark identifier: foodaudit-benchmark-450-v1.0.0.
• Full 450-record freeze status: partial public release. The public 150-record slice is frozen; the remaining 300 records are part of the internally curated benchmark and have not been cleared for unrestricted redistribution.
• Code freeze: revision baseline dated 2026-06-22. The manuscript metrics reported in the revised submission correspond to the frozen evaluation snapshot described in the paper and Online Resource 1.

Environment

Copy .env.example to .env and set credentials locally. Do not commit API keys or database passwords.

python -m pip install -r requirements.txt

Rebuild the partial release

Run scripts/build_partial_dataset.py with the source workbook and the frozen CSV exports under resources/kg_snapshot/v1/. The release builder fails when category, amount, evidence, or label-rule validation issues remain.