Car Review Sentiment, DistilBERT

• Labels: 0 negative, 1 neutral, 2 positive
• Data: Edmunds consumer car reviews, florentgbelidji/car-reviews, labels by distant supervision from the 1 to 5 star rating.
• Test set of 3,652 held-out reviews: accuracy 86.7 percent, macro-F1 0.566.
• Known limits, measured: negation and sarcasm are hard. See the stress tests in the project repo. Deploy with confidence-gated abstention.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
m = "HanfuZhao781/car-review-sentiment-distilbert"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSequenceClassification.from_pretrained(m)
x = tok("The engine is powerful but the fuel economy is terrible", return_tensors="pt")
print(model(**x).logits.softmax(-1))