A powerful 3-class sentiment analysis model fine-tuned from roberta-base on a combined corpus of 200 000+ samples spanning movie reviews, short sentences, tweets, and restaurant reviews.

🏷️ Labels

ID	Label	Description
0	Negative	Negative sentiment / opinion
1	Neutral	Neutral / mixed sentiment
2	Positive	Positive sentiment / opinion

📊 Performance

Metric	Value
Val Accuracy	0.8239
Val F1 (macro)	0.7827

🚀 Quick Usage

python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="airzipm/sentiment-analysis-roberta",
)

# Single prediction
print(classifier("This movie was absolutely amazing!"))
# [{'label': 'Positive', 'score': 0.97}]

# Batch prediction
texts = [
    "Great product, highly recommend!",
    "It was okay, nothing special.",
    "Terrible experience, waste of money.",
]
for t, r in zip(texts, classifier(texts)):
    print(f"{t[:45]:50s} → {r['label']} ({r['score']:.1%})")

🛠️ Training Details

Setting	Value
Base model	`roberta-base`
Max token length	128
Batch size	32
Learning rate	2e-5
Optimizer	AdamW + warmup
Mixed precision	FP16
Label smoothing	0.1
Class weights	Balanced

📦 Training Data

Dataset	Domain	Samples
IMDB	Movie reviews	50 000
SST-2	Short sentences	50 000
Tweet Eval	Twitter posts	50 000
Yelp Review	Business review	50 000

🖼️ Training Curves & Confusion Matrix

See training_curves.png and confusion_matrix.png in this repository.

👤 Author

Created by airzipm — Hugging Face Profile