DanishMahdi/snd_sentiment_analysis

🎬 Sindhi Movie Sentiment Analysis

Fine-tuned XLM-RoBERTa-base model for binary sentiment classification on Sindhi movie reviews.

---

📋 Model Details

Field	Details
Base model	xlm-roberta-base
Task	Binary Sentiment Classification
Language	Sindhi (sd) — Perso-Arabic / Nastaliq script
Labels	positive · negative
Dataset	DanishMahdi/snd_movies_sentiment_analysis
Training rows	~40,000 (20k positive, 20k negative)
Max token length	128

---

📊 Training Configuration

Hyperparameter	Value
Learning rate	2e-5
Batch size	16
Epochs	5 (early stopping patience = 2)
Warmup ratio	0.1
LR scheduler	cosine
Weight decay	0.01
Optimizer	AdamW
Precision	fp16 (if CUDA available)

---

🚀 Quick Start

Install dependencies

pip install transformers torch

Run inference

from transformers import pipeline

pipe = pipeline(
    "text-classification",
    model="DanishMahdi/snd_sentiment_analysis",
)

reviews = [
    "هي فلم بلڪل خراب هئي، مون کي پسند نه آئي",   # negative
    "هي فلم تمام سٺي هئي، مون کي تمام گهڻو پسند آئي",  # positive
]

for review in reviews:
    result = pipe(review)[0]
    print(f"Label: {result['label']} | Score: {result['score']:.4f}")

Output

Label: NEGATIVE | Score: 0.9873
Label: POSITIVE | Score: 0.9912

---

📁 Repository Structure

DanishMahdi/snd_sentiment_analysis/
├── config.json                    # Model config
├── model.safetensors              # Fine-tuned weights
├── tokenizer_config.json          # Tokenizer config
├── sentencepiece.bpe.model        # SentencePiece vocab
├── evaluation/
│   ├── test_metrics.json          # Accuracy, F1, Precision, Recall
│   ├── confusion_matrix.json      # Raw confusion matrix
│   ├── confusion_matrix.png       # Confusion matrix plot
│   ├── training_curves.png        # Loss & F1 over epochs
│   ├── test_metrics_bar.png       # Bar chart of metrics
│   └── classification_report.txt  # Full sklearn report
└── README.md

---

📈 Evaluation Results

See evaluation/test_metrics.json for the latest numbers. Plots are available in the evaluation/ folder.

Metric	Score
Accuracy	0.8825
F1 (weighted)	0.8825
Precision (weighted)	0.8828
Recall (weighted)	0.8825

---

🔗 Dataset

The training data is sourced from
DanishMahdi/snd_movies_sentiment_analysis

• Total rows: ~40,000
• Positive reviews: ~20,000
• Negative reviews: ~20,000
• Split: 80% train / 10% validation / 10% test

---

📝 Citation

@misc{danish2025snd,
  author    = {Danish Mahdi},
  title     = {Sindhi Movie Sentiment Analysis using XLM-RoBERTa},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/DanishMahdi/snd_sentiment_analysis},
}

---

⚖️ License

MIT — free to use for research and commercial purposes.