DavidSeyserHF/sprout-preview

plantjepa-preview

Tomato ripeness from a single fruit crop. Give it a photo plus a bounding box, get back the ripeness stage with a confidence. Small JEPA encoder plus linear head trained on pooled multi-farm boxes (BUTom-ST21, strawberry, Laboro-train) — roughly 10x smaller than SigLIP2 with far better cross-farm generality.

Use in 30 seconds

pip install -r requirements.txt

from huggingface_hub import snapshot_download
from inference import RipenessModel

folder = snapshot_download("DavidSeyserHF/sprout-preview")
model = RipenessModel.load(f"{folder}/plantjepa_preview.pt", f"{folder}/calibration.json")
print(model.predict("tomato.jpg", bbox=(x1, y1, x2, y2)))

Or from the command line:

python inference.py --checkpoint plantjepa_preview.pt --calibration calibration.json \
  --image tomato.jpg --bbox X1 Y1 X2 Y2

bbox is pixels in the original image: (xmin, ymin, xmax, ymax). The model adds a little context around the box itself, so a tight box is fine.

What you get back

Example output (stage and confidence observed; remaining values illustrative):

{
  "predicted_stage": "mature_unripe",
  "confidence": 0.5373,
  "stage_probabilities": {
    "immature": 0.01,
    "mature_unripe": 0.54,
    "ripening": 0.30,
    "ripe": 0.14,
    "overripe": 0.01
  },
  "ordinal_maturity_index_0_100": 38.2,
  "temperature": 0.8586
}

Stages are immature, mature_unripe, ripening, ripe, overripe. The maturity index is an expected rank over these stages, not a physical measurement. Confidences are temperature-scaled (see calibration.json); predictions are unchanged by it. Rule of thumb: route anything under ~0.8 confidence to human review and spot-check the rest.

How good is it?

Home rows 5–6 (5,556 crops) vs Laboro test, foreign farm (1,996 boxes):

Model	Home acc / F1	Away acc / F1
JEPA multi-source (this release)	86.84% / 77.81%	66.1% / 0.414
JEPA single-farm previous	88.01% / 79.46%	48.9% / 0.351
SigLIP2 frozen reference (~10x larger)	91.45% / 79.92%	—

Selective prediction travels even where absolute confidence does not: on Laboro, accuracy rises 0.66 → 0.89 from full coverage to the 0.95 threshold.

Limitations

• Home evidence is 2 independent greenhouse rows: point estimates, no reliable 95% CI, no equivalence claim vs SigLIP2.
• BUTom train labels are pseudo-labels; all evaluations are hand-corrected.
• The temperature in calibration.json was fit at home and is valid there only; abroad it hurts (away ECE 8.4 → 13.0). Re-fit thresholds — and ideally the temperature — per deployment instead of trusting these values.
• Calibration fixes average confidence, not worst-case confident mistakes.