sentiment_results

🧠 DistilBERT Fine-Tuned for Sentiment Analysis

A fine-tuned DistilBERT model for 3-class sentiment analysis (Negative, Neutral, Positive), trained on a large-scale sentiment dataset of ~105,000 text samples.

📋 Model Details

Model Description

• Developed by: Najaf Ali **
• Model type: Transformer-based text classification (DistilBERT)
• Language(s): English
• License: Apache 2.0
• Fine-tuned from: distilbert-base-uncased

Model Sources

• Repository: Najaf-Ali12/LLM-Hugging-Face
• Demo: Streamlit App (replace with your link)
• Training Notebook: Google Colab

---

🎯 Intended Uses

Direct Use

This model is designed for 3-class sentiment classification of English text. It can be directly used for:

• 📊 Social media monitoring – Analyzing public sentiment on tweets, posts, or comments
• 🛍️ Product review analysis – Classifying customer feedback as positive, neutral, or negative
• 💬 Customer support – Automatically prioritizing negative feedback
• 📰 Content moderation – Detecting sentiment in user-generated content
• 📈 Market research – Aggregating sentiment from survey responses or feedback forms
• 🎬 Entertainment – Analyzing movie/book reviews

Out-of-Scope Use

• Not intended for clinical, legal, or financial advice
• Not reliable for sarcasm or highly contextual humor detection
• Not validated for non-English text
• Should not be used for high-stakes decisions without human review
• Not designed for emotion detection (e.g., anger, joy, sadness) — only polarity (negative/neutral/positive)

---

⚠️ Bias, Risks, and Limitations

• Domain bias: Trained primarily on movie reviews and general text — may perform poorly on specialized domains (medical, legal, technical).
• Sarcasm detection: The model may misclassify sarcastic statements (e.g., "Oh great, another delay!").
• Short text: Very short texts (e.g., "meh", "ok") may be classified with low confidence.
• Cultural nuance: Sentiment expression varies across cultures — the model was trained on English datasets and may not capture global nuances.
• Neutral class ambiguity: The boundary between "neutral" and "slightly positive/negative" can be blurry.

Recommendation: Always use this model alongside human judgment for critical applications.

---

🚀 How to Get Started

Installation

pip install transformers torch

Quick Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "YOUR_USERNAME/YOUR_MODEL_NAME"  # Replace with your HF model ID
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Prepare input
text = "I absolutely loved this movie! The acting was brilliant."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(probabilities, dim=-1).item()

# Label mapping
id2label = {0: "NEGATIVE", 1: "NEUTRAL", 2: "POSITIVE"}
print(f"Sentiment: {id2label[predicted_class]}")
print(f"Confidence: {probabilities[0][predicted_class].item():.2%}")

Batch Prediction

texts = [
    "This product is amazing!",
    "Terrible service, never coming back.",
    "It was okay, nothing special.",
]

inputs = tokenizer(texts, return_tensors="pt", truncation=True, padding=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    preds = torch.argmax(probs, dim=-1).tolist()

for text, pred in zip(texts, preds):
    print(f"{id2label[pred]:10s} | {text}")

Using the Pipeline API

from transformers import pipeline

classifier = pipeline("text-classification", model="YOUR_USERNAME/YOUR_MODEL_NAME")

result = classifier("The food was delicious and the staff were friendly!")
print(result)
# [{'label': 'POSITIVE', 'score': 0.9876}]

---

📊 Training Details

Training Data

• Dataset: syedkhalid0/Sentiment-Analysis
• Total samples: ~105,000
  • Training: 83,989 samples
  • Validation: 10,499 samples
  • Test: 10,499 samples
• Classes: 3 (NEGATIVE = 0, NEUTRAL = 1, POSITIVE = 2)

Training Procedure

Preprocessing:

• Tokenizer: distilbert-base-uncased
• Max sequence length: 512
• Truncation: enabled
• Padding: dynamic (during collation)

Training Hyperparameters:

Hardware:

• Trained on Google Colab (NVIDIA T4 GPU recommended)

Framework:

• PyTorch + Hugging Face Transformers
• Training API: Trainer

---

📈 Evaluation Results

Test Set Performance

Per-Class Performance

💡 Tip: Run trainer.evaluate() on the test set and copy the metrics here. You can also generate a confusion matrix using scikit-learn.

Confusion Matrix

(Upload your confusion matrix image to the repo and reference it here.)

---

🧪 Example Predictions

---

🛠️ Limitations & Recommendations

• ⚠️ Sarcasm: May not detect sarcastic or ironic statements correctly.
• ⚠️ Short text: Texts under 3 words may have lower accuracy.
• ⚠️ Domain shift: Fine-tuned on general sentiment — test on your specific domain before production use.
• ✅ Recommended: For best performance, use texts between 5 and 200 words.
• ✅ Recommended: For high-stakes applications, ensemble with human review.

---

📚 Citation

If you use this model in your research or project, please cite:

@misc{distilbert-sentiment-analysis,
  author = {Najaf Ali},
  title = {DistilBERT Fine-Tuned for 3-Class Sentiment Analysis},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/YOUR_USERNAME/YOUR_MODEL_NAME}},
  note = {Fine-tuned from distilbert-base-uncased}
}

---

📞 Contact

• GitHub: @Najaf-Ali12
• Hugging Face: NajafAli01
• Issues: Please open an issue on the GitHub repository

---

🙏 Acknowledgements

• Base model: distilbert-base-uncased by Hugging Face
• Dataset: syedkhalid0/Sentiment-Analysis
• Framework: Hugging Face Transformers
• Training: Google Colab

---

📜 License

This model is released under the Apache 2.0 License. See the LICENSE file for details.

---

Last updated: 2025