roberta-finetuned-review-classifier

RoBERTa Fine-Tuned Review Classifier

This is a fine-tuned RoBERTa model built for binary sentiment classification of Amazon product reviews. The model classifies reviews into LABEL_0 (negative) or LABEL_1 (positive), and has been integrated into a Chrome extension that predicts numeric ratings based on review text.

---

Model Details

Model Description

• Model: roberta-base
• Task: Binary sentiment classification
• Labels:
  • LABEL_0: Negative review → (Rating 1–4)
  • LABEL_1: Positive review → (Rating 7–10)
• Language: English
• License: Apache 2.0
• Author: Prajjwal Chouhan
• Model repo: https://huggingface.co/prajjwal888/roberta-finetuned-review-classifier

---

Model Sources

• Base Model: roberta-base
• Fine-tuning & Deployment: Hugging Face + FastAPI (on Render)
• Demo Frontend: Chrome extension (injects rating predictions on Amazon product pages)

---

Use Cases

Direct Use

You can use this model to predict sentiment or approximate ratings of customer reviews. It's ideal for:

• Product feedback classification
• Chrome extensions or browser tools
• E-commerce dashboards

Downstream Use

Can be used in:

• Recommender systems
• Review authenticity/fraud detection
• Customer satisfaction prediction

Out-of-Scope Use

• Reviews in non-English languages
• Sarcastic or ambiguous tone detection
• Fine-grained star rating (e.g., 3 vs. 4)

---

Bias, Risks & Limitations

Bias

Model may inherit biases from training data—especially in underrepresented product categories or reviewer demographics.

Limitations

• Struggles with sarcastic or short reviews.
• Works only on English-language text.
• Predictions may be unreliable for very long reviews (truncated at 512 tokens).

Recommendations

• Do not use this model for making critical business decisions without human verification.
• Fine-tune on domain-specific reviews if required.

---

How to Get Started

from transformers import pipeline

classifier = pipeline("text-classification", model="prajjwal888/roberta-finetuned-review-classifier")
result = classifier("The product quality is fantastic. Loved it!")
print(result)

Example Output:

[{"label": "LABEL_1", "score": 0.9987}]

---

Training Details

Dataset

Custom dataset scraped and labeled from Amazon product reviews. Labeled into two categories based on review sentiment (not star ratings).

Preprocessing

• Lowercasing
• Removal of HTML and special characters
• Truncated to 512 tokens

Training Hyperparameters

---

Evaluation

Metrics

Evaluation was performed on a 20% held-out validation set from the same distribution.

---

Environmental Impact

• Hardware Used: NVIDIA T4 GPU
• Platform: Google Colab + Render
• Training Duration: ~1.5 hours
• Estimated CO₂ Emissions: ~0.3 kg (based on ML CO2 Impact Calculator)

---

Technical Specifications

• Model Type: Transformer Encoder (RoBERTa)
• Architecture: 12-layer, 768-hidden, 12-heads, ~125M parameters
• Framework: PyTorch (via transformers)

---

Citation

@misc{prajjwal888_review_classifier_2024, title={RoBERTa Fine-Tuned Review Classifier}, author={Prajjwal Chouhan}, year={2024}, howpublished={\url{https://huggingface.co/prajjwal888/roberta-finetuned-review-classifier}}, }

---

Contact

• GitHub: @prajjwal888
• LinkedIn: Prajjwal Chouhan

---

Acknowledgments

Thanks to the Hugging Face community and the creators of roberta-base. This project is inspired by practical applications of NLP in e-commerce.