diyaolhaqq's picture
Create eval/SMOKE_TESTS.md
8bd7c2b verified
|
Raw
History Blame Contribute Delete
2.22 kB

Functional smoke-test record

Evidence status: Research-preview maize screening model evaluated on documented development datasets and limited field samples.

These recorded checks used non-redistributed local reference images to verify model loading, tensor shape, preprocessing, and routing behavior. They are implementation evidence only—not an accuracy benchmark, a leakage-safe test set, an agronomist-confirmed diagnosis, or field validation.

Test Expected pathway Recorded status
Healthy maize development reference Stage 1 → healthy Pass — healthy 0.976533
Indoor/non-maize development reference Stage 1 → not maize/unclear Pass — not_maize_or_unclear 0.967990
Limited maize field sample; disease label unconfirmed Stage 1 → Stage 2 Pass — Stage 1 diseased 0.980578; Stage 2 screened northern_leaf_blight 0.911019

Verification was recorded with Python 3.12, TensorFlow 2.20.0, and the packaged FP16 files. The reference images are not included in this staging directory, so the recorded image-level outputs are not independently reproducible from this package alone.

The limited field sample verifies only that the routing and symptom-crop code can run end to end. Its disease label was not independently confirmed by an agronomist, and the predicted class is not accuracy evidence. Field validation is incomplete.

Separately, a retrospective audit of the original v1.4 metadata.csv found 80,313 records, 68,786 source + group IDs, and 2,205 groups crossing train/validation/test. The historical 0.992/0.993 development metrics are not leakage-safe release results and must not be presented as public or field performance. The detailed evidence is in development-data-audit.json.

corrected-split-plan.json records a deterministic group-disjoint assignment of 57,914 train, 11,107 validation, and 11,292 test records, with zero cross-split groups and records. That is a repair plan—not new training or evaluation evidence. The packaged models have not been retrained or evaluated on the corrected assignment, and the independent Ghana field holdout is incomplete.