geraldmc/plantvillage-tiny

PlantVillage (tiny)

This is a debug-grade subset, not a faithful subsample for analysis. 50 images per class drawn from the full PlantVillage dataset is too few to represent class-level visual diversity. Use it for iterating on training-loop code, smoke-testing pipelines, or any situation where you want the data structure but not the data scale. For actual classifier training or evaluation, use geraldmc/plantvillage-full.

What's in this dataset

A stratified subsample of geraldmc/plantvillage-full — 50 images per class × 38 classes = approximately 1,900 images total. The schema is identical to the full variant; only the row count differs. Built so that pipelines depending on the full dataset's structure (column layout, image format, train/test split shape) can be developed and tested without the ~855 MB download every iteration.

How to use it

from datasets import load_dataset

ds = load_dataset("geraldmc/plantvillage-tiny", revision="v0.1.0", split="train")

For users of the irilab2026 library:

import irilab2026 as iri
metadata, images = iri.load_plantvillage(variant="tiny")

Data fields

Identical to geraldmc/plantvillage-full. See that dataset's card for the field-by-field schema.

Build provenance

Built by the same script (scripts/build_pv_full_hf.py) and from the same upstream commit as plantvillage-full. The additional step is stratified subsampling:

• Random seed: 20260519 (fixed for reproducibility).
• 50 images per class, sampled without replacement, per class independently.
• The train/test split column is inherited from the full variant's pre-split assignment, not re-shuffled. This preserves leaf-grouping within the subsample, but means the train/test ratio in the tiny variant may drift slightly from 80/20.

Known caveats

Sample size is too small for analytical claims. Per-class accuracy on ~10 held-out test images per class is noisy enough that any reported number swings with the seed. Use the full variant for any result you intend to report.

Same upstream caveats apply. The image count discrepancy and the Tomato___Target_Spot leafmap issue documented for the full variant also apply here, in subsampled form. See the full variant's dataset card for details.

License

CC0 1.0 (Public Domain Dedication), inherited from the upstream PlantVillage release.

Citation

Same as the full variant. See geraldmc/plantvillage-full's dataset card for the BibTeX entry.