Kapilydv6/my-first-sentiment-model

My First Sentiment Analysis Model 🎬

A fine-tuned DistilBERT model for classifying movie reviews as POSITIVE or NEGATIVE.

How to Use

from transformers import pipeline

classifier = pipeline("sentiment-analysis", model="Kapilydv6/my-first-sentiment-model")

result = classifier("This movie was amazing!")
print(result)
# [{'label': 'POSITIVE', 'score': 0.99}]

Training Details

• Base model: distilbert-base-uncased
• Dataset: IMDB (2000 training samples)
• Epochs: 3
• Framework: PyTorch + Hugging Face Transformers

Limitations

This is a beginner project trained on a small subset of IMDB reviews. For production use, train on the full dataset with more epochs.

Training Results Explained

📌 What Happened Step by Step

1. Dataset Loaded

• Downloaded the full IMDB dataset (25,000 train + 25,000 test reviews)
• Used only:
  • 2,000 samples for training
  • 500 samples for testing
• This was done as per configuration

---

2. Model Loaded — "MISSING / UNEXPECTED" Warnings

These warnings are normal and expected:

• UNEXPECTED keys (vocab layers)

  • DistilBERT was originally trained for masked language modeling
  • These layers are not needed for classification

• MISSING keys (classifier, pre_classifier)

  • These are new layers added for the sentiment task
  • They are randomly initialized and learned during training

---

3. Training (3 Epochs — ~1h 46min on CPU)

Epoch	Train Loss	Eval Accuracy	Eval F1
1	~0.51	85.0%	84.96%
2	~0.26	86.6%	86.55%
3	~0.16	87.4%	87.39%

✅ Loss decreased (0.60 → 0.15) → Model is learning
✅ Accuracy increased each epoch → Model is improving
✅ Best model (Epoch 3) was automatically selected and saved

---

4. Final Accuracy

• 87.4% accuracy
• Strong performance considering only 2,000 training samples
• Using the full dataset could reach ~93%+ accuracy

---

5. Test Predictions

Review	Prediction	Confidence
"Absolutely fantastic!"	POSITIVE	97.9%
"Terrible waste of time"	NEGATIVE	96.7%
"It was okay, nothing special"	NEGATIVE	75.3%

💡 Insight:

• The third review is ambiguous
• A human might label it neutral
• The model predicts NEGATIVE with lower confidence, which is reasonable

---

⚠️ Warnings to Ignore

• Symlinks warning

  • Harmless Windows limitation
  • Only affects disk usage

• LayerNorm gamma/beta → weight/bias

  • Due to PyTorch naming changes
  • No impact on results

---

✅ Summary

• Model trained successfully
• Performance improved across epochs
• Achieved strong accuracy with limited data
• Predictions are confident and reasonable
• Warnings are expected and safe to ignore

---