MrZaper/LiteModel

MrZaper/LiteModel is a lightweight sentence-transformers model fine-tuned for semantic search and retrieval of academic articles in cybersecurity.
It maps queries and article phrases into a 384-dimensional dense vector space for similarity search, clustering, and semantic matching.

This model is specifically trained for the journal: Cybersecurity: Education, Science, Technology
Website: https://csecurity.kubg.edu.ua

What does it do?

Given a query in English, Ukrainian, or any other language, the model:

• Translates the query to English (using Google Translate).
• Encodes the query into a dense embedding using Sentence-BERT.
• Computes cosine similarity between the query embedding and precomputed article embeddings.
• Returns the top unique article codes with highest similarity scores.

Returned article codes can be viewed at:

https://csecurity.kubg.edu.ua/index.php/journal/article/view/{CODE}

For example:

560 → https://csecurity.kubg.edu.ua/index.php/journal/article/view/560

Model Files

The repository includes:

• LiteModel – SBERT-based semantic encoder
• sbert_embeddings.npy – Precomputed embeddings for articles
• sbert_labels.pkl – Corresponding article codes (e.g., 560, 532)

Usage (Sentence-Transformers)

Install the required package:

pip install -U sentence-transformers deep-translator huggingface-hub scikit-learn

Example usage:

from sentence_transformers import SentenceTransformer
import numpy as np
import pickle
from huggingface_hub import snapshot_download
from deep_translator import GoogleTranslator
import os
from sklearn.metrics.pairwise import cosine_similarity

# Load model and data from Hugging Face
model_name = 'MrZaper/LiteModel'
model_dir = snapshot_download(repo_id=model_name)

# Load SBERT model
sbert_model = SentenceTransformer(model_dir)

# Load precomputed article embeddings
embeddings = np.load(os.path.join(model_dir, "sbert_embeddings.npy"))

# Load article codes (labels)
with open(os.path.join(model_dir, "sbert_labels.pkl"), 'rb') as f:
    labels = pickle.load(f)

def preprocess_query(query: str) -> str:
    """Translate the query to English using Google Translate."""
    try:
        return GoogleTranslator(source="auto", target="en").translate(query)
    except Exception as e:
        print(f"Translation error: {e}")
        return query

def predict_semantic(query, model, embeddings, labels, top_n=5):
    """Find top-N most semantically similar unique article codes."""
    query_emb = model.encode([preprocess_query(query)])
    similarities = cosine_similarity(query_emb, embeddings)[0]

    seen_keys = set()
    results = []

    # Sort results by similarity (descending)
    sorted_indices = np.argsort(similarities)[::-1]

    for idx in sorted_indices:
        label = labels[idx]
        sim = similarities[idx]

        if label not in seen_keys:
            seen_keys.add(label)
            results.append({
                "article_code": label,
                "similarity": float(sim)
            })
            print(f"📄 Article {label} – similarity: {sim * 100:.2f}%")

        if len(results) >= top_n:
            break

    return results

# Example query
query = "sql injection in websites"
results = predict_semantic(query, sbert_model, embeddings, labels)

print("\nTop article codes:")
for res in results:
    print(f"Article {res['article_code']} – similarity: {res['similarity']*100:.2f}%")

Example Output

📄 Article 560 – similarity: 92.15%
📄 Article 532 – similarity: 89.34%
📄 Article 475 – similarity: 85.22%

Corresponding links:

https://csecurity.kubg.edu.ua/index.php/journal/article/view/560
https://csecurity.kubg.edu.ua/index.php/journal/article/view/532
https://csecurity.kubg.edu.ua/index.php/journal/article/view/475