SentiLens — Fine-Tuned DistilBERT

Overview

SentiLens is a binary sentiment classification model fine-tuned on the IMDB movie review dataset. It predicts whether a movie review expresses positive or negative sentiment.

This project explores transformer fine-tuning, model evaluation, and the comparison of a specialised classifier with large language model prompting.

Model details

• Base model: distilbert/distilbert-base-uncased
• Task: Binary sentiment classification
• Dataset: stanfordnlp/imdb
• Training epochs: 2
• Maximum sequence length: 256 tokens
• Training batch size: 8
• Learning rate: 2e-5
• Labels: 0 = NEGATIVE, 1 = POSITIVE

Evaluation

Evaluation on the 25,000-example IMDB test set produced:

Confusion matrix (rows: actual; columns: predicted):

These results apply to the IMDB test set and should not be interpreted as performance guarantees on other types of text.

Intended use

• Educational sentiment analysis
• Movie review classification
• Demonstrations of transformer fine-tuning

Limitations

• The model is specialised for English-language movie reviews.
• Performance may differ on social media, sarcasm, mixed sentiment, other languages, or domains outside movie reviews.
• A sentiment label does not explain every nuance of a review.
• This model is not intended for high-stakes decisions.

Training and evaluation

The model was fine-tuned using Hugging Face Transformers and PyTorch in Google Colab. Evaluation metrics were calculated on the dataset's held-out test split.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="princy-21/sentilens-distilbert"
)

print(classifier("The acting was excellent and the story was moving!"))

Acknowledgements

• Base model: DistilBERT, originally developed by the DistilBERT authors.
• Dataset: IMDB movie reviews distributed through Hugging Face Datasets.

Review the original model and dataset licenses and attribution requirements before redistributing this model.

Author

Amal Princy J — student AI/ML project.