sarkarghya/yelp-fake-review-detector

Yelp Fake Review Detector

DeBERTa-v3-base fine-tuned on YelpZip filter-status proxy labels.

The labels are proxy ground truth from the YelpZip dataset, not manually verified fraud labels. Use this model to identify reviews that match Yelp-filtered suspicious patterns, not to make an unsupported claim that a person or review is fraudulent.

Labels

• recommended: Yelp-filtered recommended-like review
• yelp_filtered_suspicious: Yelp-filtered suspicious pattern

Evaluation

The model was trained on 483,363 reviews, validated on 60,142 reviews, and tested on 60,496 reviews. The full metrics are included in metrics.json.

At the tuned threshold in threshold.json:

• Threshold: 0.69
• Accuracy: 0.8057
• F1: 0.4033
• ROC AUC: 0.7855
• PR AUC: 0.3706

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="sarkarghya/yelp-fake-review-detector",
)
print(classifier("The service was friendly and the meal was fresh."))

For a binary decision matching the reported evaluation, compare the probability for yelp_filtered_suspicious with the threshold in threshold.json.