Fine-tuned BERT for Sentiment Analysis on SST-2
This model is a fine-tuned version of BERT (bert-base-uncased) specifically designed for binary sentiment classification of English text, achieving state-of-the-art performance on the Stanford Sentiment Treebank v2 (SST-2) benchmark.
Model Description
The model was created to demonstrate the practical application of transfer learning in Natural Language Processing (NLP). While the base BERT model has a deep understanding of general English language structure, it was not originally trained to detect sentiment. This fine-tuning process adapts BERT's powerful contextual embeddings to the specialized task of determining whether a given sentence expresses a positive or negative opinion.
Key Technical Details
- Architecture: BERT-base-uncased with a sequence classification head (2 output neurons).
- Training Approach: The pre-trained BERT layers were gently tuned while the newly added classification layer was trained from scratch over 3 epochs.
- Framework: PyTorch with the Hugging Face Transformers library.
Intended Use & Limitations
✅ Intended Use
This model is optimal for classifying the sentiment of short English texts, particularly:
- Movie or product reviews
- Social media posts (opinions)
- Customer feedback snippets
⚠️ Limitations
- Domain Specificity: Performance may degrade on texts far outside the movie review domain (e.g., technical, financial, or medical jargon).
- Binary Scope: It is designed for positive/negative classification and does not detect neutral sentiment or more complex emotions.
- Language: Works only with English text.
Training Data
The model was fine-tuned on the Stanford Sentiment Treebank v2 (SST-2) dataset from the GLUE benchmark.
| Dataset Split | Number of Examples |
|---|
| Training | 67,349 |
| Validation | 872 |
| Test | 1,821 |
Example from the dataset:
- Sentence: "contains no wit , only labored gags"
- Label:
0 (Negative)
Training Procedure & Hyperparameters
The model was trained for 3 epochs using the following configuration:
| Hyperparameter | Value |
|---|
| Learning Rate | 2e-5 |
| Batch Size | 16 |
| Optimizer | AdamW |
| Weight Decay | 0.01 |
| Warmup Steps | 0 |
| Max Sequence Length | 128 |
The training leveraged the Hugging Face Trainer API for efficient optimization and evaluation.
Evaluation Results
The model's performance was evaluated on the SST-2 validation set, yielding the following metrics:
📊 Overall Performance
| Metric | Score |
|---|
| Accuracy | 92.55% |
| F1-Score (Macro Avg) | 0.93 |
| Precision (Negative) | 0.93 |
| Recall (Positive) | 0.94 |
📈 Training Progress
| Epoch | Training Loss | Validation Loss | Validation Accuracy |
|---|
| 1 | 0.1760 | 0.2400 | 92.43% |
| 2 | 0.1240 | 0.3320 | 91.63% |
| 3 | 0.0704 | 0.3400 | 92.55% |
Confusion Matrix (Validation Set, n=872)
| Predicted Negative | Predicted Positive |
|---|
| Actual Negative | 391 (TN) | 37 (FP) |
| Actual Positive | 27 (FN) | 417 (TP) |
Live Inference Examples
The model correctly classifies clear examples and shows nuanced understanding of ambiguous text:
| Input Sentence | Predicted Label | Confidence (Negative, Positive) |
|---|
| "The movie was fantastic!" | Positive | [0.0002, 0.9998] |
| "I hated every minute of this film." | Negative | [0.9994, 0.0006] |
| "It was okay, nothing special." | Positive | [0.4308, 0.5692] |
Note: The third example shows low confidence, appropriately reflecting the neutral sentiment of the input.
Conclusion
This project successfully demonstrates how transfer learning with a foundation model like BERT can efficiently create a high-performance, specialized classifier. With minimal training time and data, the fine-tuned model achieves competitive results on a standard NLP benchmark, making it suitable for real-world sentiment analysis applications.
Model card generated using best practices from the Hugging Face Model Card Guidebook. 
