geraldmc/plantdoc-tiny

PlantDoc — tiny variant

Warning: this is a debug-grade subset, not a faithful subsample for analysis. Use it for test suites, smoke tests, and notebook iteration. For substantive work, use geraldmc/plantdoc-full.

A 164-image stratified subsample of geraldmc/plantdoc-full, built to support fast iteration. Loading is roughly a 50 MB download instead of ~950 MB, and a pass over the dataset takes seconds rather than minutes.

Quick start

from datasets import load_dataset

ds = load_dataset("geraldmc/plantdoc-tiny", revision="v0.1.0", split="train")
print(len(ds))    # 164

Inside the iResearch Institute 2026 Virtual Lab:

import irilab2026 as iri

metadata_df, hf_dataset = iri.load_plantdoc(variant="tiny")

What's in this dataset

• 164 images stratified across 28 classes, 83 train + 81 test
• Schema identical to plantdoc-full — same 8 columns (image + 7 metadata), same values for any image that appears in both
• Subsample rule: min(3, available) per class per split, with a fixed seed

For every non-orphan class, this means 3 train + 3 test = 6 images. The orphan class (Tomato two spotted spider mites leaf) appears with its 2 train images and 0 test images, exactly as in the full variant.

Schema, normalization, caveats

See the full variant's dataset card for:

• Column descriptions and types
• Normalization rules for host and disease
• The orphan-class caveat
• Filename quirks
• Citation and license details

The schema is identical between the two variants; duplicating it here invites drift.

Subsample parameters

• Per-class budget: min(3, available) per (class, split)
• Seed: 42 (passed as random_state to pandas .sample())
• Source: geraldmc/plantdoc-full at revision v0.1.0

The subsample is fully deterministic given these three parameters. Re-running the build script with the same source and seed produces byte-identical metadata. Image bytes also match exactly because the same upstream files are referenced.

What this dataset is NOT for

• Training a classifier. 3 training images per class is far too few to learn anything; this isn't a "small training set" — it's a fixture for code paths.
• Measuring per-class accuracy. Per-class test set sizes here are 1–3 images; any "accuracy" you compute is dominated by sampling noise.
• Comparing model architectures. Any signal would be noise at this scale.
• Reporting results in a paper. Whatever you find on the tiny variant is not a finding about PlantDoc — it's a finding about 164 images that happen to live under this repo ID.

What this dataset IS for

• Test suite use. Fast assertion that loader code, Dataset wrappers, and training pipelines run end-to-end without surprise.
• Notebook iteration. Debugging a preprocessing pipeline or augmentation strategy without paying for the full dataset's load time.
• CI runs. Tests that exercise loader behavior can use this variant on every push without burning minutes on downloads.

License

CC BY 4.0, matching the full variant and the upstream dataset.

Citation

Cite the upstream dataset (Singh et al. 2020) and reference this tiny variant as geraldmc/plantdoc-tiny @ v0.1.0. Full BibTeX in the full variant's dataset card.