What the code actually asks
Eşya Bul’s public verifier chooses candidate product groups from the listing title or an existing category tag. For a refrigerator, it compares a positive description of the complete appliance with alternatives such as a shelf, motor, handle or repair service. The inspected version uses CLIP ViT-B/32 and applies softmax to the image logits for that particular set of candidate descriptions.
The question is not whether the seller is honest. It is which of the supplied descriptions best matches the photograph. That can be a useful starting point for filtering spare parts advertised under an appliance title. Working condition, ownership, availability, delivery and payment safety remain separate properties that these inputs cannot establish.
Why a percentage-shaped number can mislead
The implementation retains a listing when its positive score exceeds 0.85. Storing that number under confidence_score does not demonstrate 85 percent accuracy. The softmax denominator depends on the candidate texts supplied to the comparison. Changing those candidates can change a photograph’s score without retraining the model.
This small calculation demonstrates only that normalization effect. It is not a real CLIP output or an Eşya Bul benchmark. The positive logit stays fixed, but adding a strong alternative reduces its share. A threshold becomes operationally useful only after evaluation with a fixed candidate design and labeled examples.
from math import exp
def positive_share(logits):
top = max(logits)
weights = [exp(value - top) for value in logits]
return weights[0] / sum(weights)
print(round(positive_share([4, 1, 0]), 3)) # 0.936
print(round(positive_share([4, 1, 0, 4]), 3)) # 0.484Three blind spots in this implementation
First, the title and previous category tags influence which comparisons are attempted. An incorrect category can cause the model to answer the wrong question. Second, this function considers an individual image, without jointly interpreting other listing photos or damage described in the text. Third, a download failure and an unsuccessful classification both disappear from the resulting list.
The number of missing listings is therefore not a measure of filtering quality. Unreachable images, titles with no matching category and low scores need distinct outcome reasons. OpenAI’s model card likewise emphasizes evaluation in the intended context and sensitivity to category design, with additional limitations around non-English usage and fine-grained distinctions.
Define the labels before running the experiment
Begin with a narrow question: does the photograph show a complete refrigerator or a replacement part? Do not report that experiment as an evaluation of whether a listing is safe. Keep ambiguous collages, screenshots and sealed packages in a separate group. This makes difficult cases visible rather than forcing uncertain human judgments into an apparently clean binary dataset.
Record the reason alongside each human label. If a photograph contains both a complete appliance and a removed part, understanding why two reviewers disagree can guide the candidate descriptions. Preserve disagreements in an error-review list instead of hiding them; they help define the question the next experiment should resolve.
| Example group | Human label | Purpose |
|---|---|---|
| Complete product | Match | Detect good listings incorrectly rejected |
| Part or repair advertisement | Mismatch | Detect false accepts |
| Collage or unclear image | Uncertain | Inspect difficult cases |
| File cannot be downloaded | Technical error | Separate network and model failures |
Choose a threshold from observed outcomes
Separate development examples from a held-out evaluation set, keeping near-duplicate photographs of the same listing on one side. Report false accepts and false rejects for several thresholds, broken down by category. Missing a useful listing may dominate the cost when people review a shortlist; false acceptance may matter more when the output triggers an automated action. The threshold should reflect that product decision.
This article does not announce a new benchmark. The defensible description supported by the inspected code is preliminary filtering for visual category consistency. Narrowing the claim helps users understand what the model contributes and which judgments still require evidence beyond an image similarity score.