Kafwain Produce Freshness (ViT)

An image classifier that looks at a photo of fresh produce and predicts what it is and whether it looks fresh or spoiled. It is a fine-tuned version of google/vit-base-patch16-224, trained on v3loc1ty/kafwain-produce-freshness.

It reaches 99.68% accuracy on its evaluation set. That set comes from the same kind of images as the training data, so expect lower accuracy on real-world photos (see Limitations).

The model is a prototype for Kafwain, a student startup in Qatar that connects local farms with restaurants through one order, one chilled delivery and one invoice. In Kafwain, the model's job is to help staff check produce when farms drop off orders, flagging items that may be spoiled so a person can look at them.

Model details

Labels

The model predicts one of these labels. Each label combines a produce type with a freshness state.

[FILL: copy the label list from id2label in this repo's config.json.]

Produce types not in this list are not supported. The model will still return one of the labels above, often with high confidence (see Limitations).

Intended uses

Primary use: an assistant for the pickup check. At Kafwain's pickup point, staff photograph produce as farms drop it off. The model flags items that may be spoiled, and a person makes the final decision to accept or reject them.

Other uses:

• Demonstrating Kafwain's planned freshness check in pitches and prototypes.
• Learning and experimenting with image classification on produce.

Intended users: Kafwain operations staff (through the Kafwain app) and developers building or testing the feature.

Out-of-scope uses

Do not use this model for:

• Food safety decisions. "Looks fresh" doesn't mean "safe to eat." The model can't detect contamination, pesticides, bacteria or mould inside produce.
• Automatically rejecting a farm's produce, or setting prices, with no person checking. Errors would cost farms money unfairly.
• Quality grading (e.g. grade A vs grade B). It wasn't trained for that.
• Produce types not in the label list, or images that aren't produce.
• Consumer, medical or health advice.

How to use

Quick start with a pipeline:

from transformers import pipeline

classifier = pipeline("image-classification", model="v3loc1ty/kafwain-produce-freshness-vit")
print(classifier("produce_photo.jpg"))

Recommended use, with a confidence threshold so uncertain photos go to a person:

import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForImageClassification

model_id = "v3loc1ty/kafwain-produce-freshness-vit"
processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModelForImageClassification.from_pretrained(model_id)

image = Image.open("produce_photo.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    probs = model(**inputs).logits.softmax(dim=-1)[0]

score, idx = probs.max(dim=-1)
label = model.config.id2label[idx.item()]

THRESHOLD = 0.80  # starting value; tune it on your own photos
if score.item() < THRESHOLD:
    print("Unsure: send to a person to check")
else:
    print(f"{label} ({score.item():.1%})")

Photo tips: one item or a close-up of a few items, filling most of the frame, in good even light, on a plain background.

Training data

Trained and evaluated on v3loc1ty/kafwain-produce-freshness.

Training procedure

Fine-tuned with the Hugging Face Trainer.

Framework versions: Transformers 5.22.2 · PyTorch 2.15.0+cu132 · Datasets 5.1.0 · Tokenizers 0.23.3

Evaluation

Results on the evaluation set

Final: loss 0.0123, accuracy 99.68% on [FILL: evaluation split name and number of images].

Per-label results

[FILL: precision, recall and F1 for each label, and a confusion matrix. Overall accuracy can hide a weak label, especially a rare one.]

Real-world test (planned)

The model hasn't yet been tested on Kafwain's own photos taken at a pickup point. This will be added after the pilot. Until then, treat 99.68% as an upper bound, not the expected real-world accuracy.

Limitations

• The evaluation score is probably optimistic. The evaluation images come from the same sources and conditions as the training images. Public freshness datasets also often contain near-identical images (augmented copies, or several shots of the same item). If a random split put copies of one item in both training and evaluation, accuracy is inflated.
• Real photos look different. Public datasets mostly show single items on plain backgrounds. Crates, bags, mixed items, shadows, glare, wet produce and cluttered backgrounds will reduce accuracy.
• Limited produce types. Only the labels listed above are supported. Many items Kafwain sells, such as leafy greens and herbs, may not be covered.
• It always picks a label. Shown something it doesn't know, the model still returns one of its labels, often with high confidence. Use the confidence threshold and send uncertain photos to a person.
• Only what's visible. It judges the surface in the photo. It can't see the other side of an item, the inside of a crate, or spoilage that hasn't shown on the skin yet.
• Two states only. Fresh vs spoiled misses the stages in between, such as slightly bruised or near the end of shelf life.
• Not tested on local produce. The training data likely comes from outside Qatar. Local varieties, sizes and colors may differ.

Bias, risks and recommendations

• Unfair rejections. A wrong "spoiled" prediction could cost a farm money. Keep a person in charge of every accept or reject decision, and let farms dispute a decision.
• Over-trust. A high score can look certain when it isn't. Show the score and the photo to staff, not just the label.
• Before real use:
  1. Collect and label photos from the actual pickup point.
  2. Re-split the data so that photos of the same item never appear in both training and evaluation.
  3. Report per-label results and tune the confidence threshold on those photos.
  4. Add the produce types Kafwain actually sells.

Citation

Base model:

@misc{dosovitskiy2020image,
  title         = {An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale},
  author        = {Alexey Dosovitskiy and Lucas Beyer and Alexander Kolesnikov and Dirk Weissenborn and Xiaohua Zhai and Thomas Unterthiner and Mostafa Dehghani and Matthias Minderer and Georg Heigold and Sylvain Gelly and Jakob Uszkoreit and Neil Houlsby},
  year          = {2020},
  eprint        = {2010.11929},
  archivePrefix = {arXiv}
}

Source datasets: [FILL: credit and cite each dataset used to build v3loc1ty/kafwain-produce-freshness, following its license.]

Model card contact

The Kafwain team: [FILL: contact email or link]