This model classifies Uzbek text messages as either spam or normal (ham). It was fine-tuned on a synthetic dataset of 2,000 Uzbek messages covering common spam patterns found in Telegram and messaging platforms.

Property	Value
Base Model	xlm-roberta-base
Language	Uzbek (Latin & Cyrillic scripts)
Task	Binary Text Classification
Labels	`spam`, `normal`

Intended Use

Primary Use Cases

Spam filtering for Uzbek Telegram bots
Content moderation for Uzbek social platforms
Message classification in Uzbek chat applications

Out-of-Scope Use

Languages other than Uzbek
Long-form document classification
Detecting other types of harmful content (hate speech, etc.)

How to Use

Quick Start

python

from transformers import pipeline

# Load the model
classifier = pipeline("text-classification", model="sukhrobnurali/uzbek-spam-detector")

# Classify messages
result = classifier("Salom! Bugun uchrashuvga kela olasanmi?")
print(result)
# [{'label': 'normal', 'score': 0.98}]

result = classifier("TEZKOR KREDIT! 50% chegirma! Bosing: example.com")
print(result)
# [{'label': 'spam', 'score': 0.99}]

Batch Classification

python

messages = [
    "Rahmat katta yordam uchun!",
    "Tabriklaymiz! Siz 1000$ yutdingiz!",
    "Kecha juda charchab uyga keldim",
]

results = classifier(messages)
for msg, res in zip(messages, results):
    print(f"{res['label']}: {msg[:40]}...")

Using with PyTorch

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("sukhrobnurali/uzbek-spam-detector")
model = AutoModelForSequenceClassification.from_pretrained("sukhrobnurali/uzbek-spam-detector")

text = "Bizning kanalga obuna bo'ling!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=-1)

label = model.config.id2label[prediction.item()]
print(f"Prediction: {label}")

Training Details

Training Data

The model was trained on sukhrobnurali/uzbek_spam_dataset, a synthetic dataset of 2,000 Uzbek messages:

Split	Samples	Spam	Normal
Train	1,800	~900	~900
Test	200	~100	~100

Spam Categories Covered

Aggressive advertising and promotions
Get-rich-quick schemes
Unsolicited loan/credit offers
Fake prize/giveaway announcements
Clickbait messages
Channel/group promotion spam

Training Procedure

Parameter	Value
Base model	`xlm-roberta-base`
Epochs	3
Batch size	16
Learning rate	2e-5
Weight decay	0.01
Warmup ratio	0.1
Max sequence length	128
Optimizer	AdamW
Precision	FP16

Evaluation Results

Performance on the held-out test set (200 samples):

Metric	Score
Accuracy	100.0%
F1 Score	100.0%
Precision	100.0%
Recall	100.0%

Classification Report

              precision    recall  f1-score   support

      normal       1.00      1.00      1.00       106
        spam       1.00      1.00      1.00        94

    accuracy                           1.00       200
   macro avg       1.00      1.00      1.00       200
weighted avg       1.00      1.00      1.00       200

Note: The perfect scores are due to the synthetic nature of the training data, where spam and normal messages have distinct, learnable patterns. Real-world performance may vary with organic messages that have more subtle spam indicators.

Limitations

Synthetic Data: The model was trained on AI-generated messages, which may not capture all real-world spam patterns.
Domain Specific: Optimized for Telegram-style short messages. Performance may vary on:
- Long-form content
- Formal documents
- Other messaging platforms
Language Coverage: Primarily tested on Uzbek. May have unpredictable behavior on:
- Code-mixed Uzbek-Russian text
- Heavy use of transliteration
Evolving Spam: Spam tactics change over time. The model may need retraining to catch new patterns.

Ethical Considerations

False Positives: The model may incorrectly flag legitimate messages as spam. Always provide users a way to report misclassifications.
Bias: Synthetic training data may contain biases from the generation model.
Privacy: This model processes text locally and does not store or transmit user messages.

Citation

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

bibtex

@misc{uzbek-spam-detector,
  author = {Sukhrob Nurali},
  title = {Uzbek Spam Detector: Fine-tuned XLM-RoBERTa for Uzbek Spam Classification},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/sukhrobnurali/uzbek-spam-detector}
}

License

This model is released under the MIT License.

sukhrobnurali/uzbek-spam-detector

Uzbek Spam Detector

Model Description

Intended Use

Primary Use Cases

Out-of-Scope Use

How to Use

Quick Start

Batch Classification

Using with PyTorch

Training Details

Training Data

Spam Categories Covered

Training Procedure

Evaluation Results

Classification Report

Limitations

Ethical Considerations

Citation

Links

License

sukhrobnurali/
uzbek-spam-detector