A BERT-base model fine-tuned for 5-class review-rating classification — give it a free-text review and it predicts a rating from 1 to 5 stars (the yelp_review_full label scheme).

Built by Amit Aminov · amitaminov.github.io

Model at a glance


Architecture	`BertForSequenceClassification` — BERT-base (12 layers, hidden 768)
Task	Text classification — review rating
Classes	5 (`LABEL_0`…`LABEL_4` = 1★…5★, in `yelp_review_full` order)
Max sequence length	512 tokens
Language	English
Fine-tuned from	`bert-base-cased`

Usage

python

from transformers import pipeline

clf = pipeline("text-classification", model="AmitAminov/yelp_review_classifier")
clf("Incredible food and the service was so warm — we'll be back every week.")
# -> [{'label': 'LABEL_4', 'score': ...}]   # LABEL_4 = 5 stars

The raw labels map to stars as LABEL_0 → 1★, LABEL_1 → 2★, …, LABEL_4 → 5★.

Load the model directly

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tok = AutoTokenizer.from_pretrained("AmitAminov/yelp_review_classifier")
model = AutoModelForSequenceClassification.from_pretrained("AmitAminov/yelp_review_classifier")

Intended use & limitations

Intended for rating-style sentiment on English review text (restaurants, businesses, and products with a similar 1–5 register).
Labels are the raw LABEL_0…LABEL_4 from training — interpret them as 1–5 stars.
Very short snippets, non-review language, or other languages will be less reliable, and — as with any sentiment model — predictions can reflect biases in the training data.

Training

Fine-tuned with the 🤗 Transformers Trainer on the yelp_review_full dataset (5-star reviews). Held-out evaluation metrics were not recorded in this repository; if you need them, re-run evaluation on the yelp_review_full test split.

_{Part of Amit Aminov's public work.}