Upload folder using huggingface_hub

README.md

@@ -6,9 +6,9 @@ tags: [vision, ripeness, tomato, jepa, 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 calibrated confidence. Small JEPA encoder
-plus linear head — roughly 10x smaller than SigLIP2 at nearly the same
-class-averaged quality.
+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
 
@@ -63,18 +63,23 @@ review and spot-check the rest.
 
 ## How good is it?
 
-Held-out tomato rows 5 and 6 (5,556 crops):
+Home rows 5–6 (5,556 crops) vs Laboro test, foreign farm (1,996 boxes):
 
-| Model | Accuracy | Macro-F1 |
+| Model | Home acc / F1 | Away acc / F1 |
 |---|---|---|
-| JEPA adapted (this release) | 88.01% | 79.46% |
-| SigLIP2 frozen reference (~10x larger) | 91.45% | 79.92% |
+| 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% | — |
 
-Calibration cut expected calibration error from 4.04% to 1.86% on the same rows.
+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
 
-- Only 2 independent greenhouse test rows: point estimates, no reliable 95%
-  CI, no equivalence claim vs SigLIP2.
-- Trained on pseudo-labels; evaluated on hand-corrected labels.
+- 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.