KazEmbed-V5: Kazakh Embedding Model for RAG

🏆 Best BASE-size embedding model for Kazakh language retrieval tasks

Model Description

KazEmbed-V5 is a fine-tuned version of multilingual-e5-base optimized for Kazakh language retrieval and RAG (Retrieval-Augmented Generation) applications.

Key Features

• 🇰🇿 Specialized for Kazakh: Fine-tuned on 61,255 Kazakh text pairs
• 📈 +2.1% MRR improvement over multilingual-e5-base
• ⚡ Efficient: 278M parameters (base-size model)
• 🔍 RAG-optimized: Trained specifically for retrieval tasks

Benchmark Results

Evaluated on KazQAD test set with TF-IDF hard negatives (100 candidates per query)

Usage

Installation

pip install sentence-transformers

Basic Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('YOUR_USERNAME/kazembed-v5')

# For queries (questions)
query = "query: Қазақстанның астанасы қай қала?"
query_embedding = model.encode(query)

# For passages (documents)
passage = "passage: Астана — Қазақстан Республикасының астанасы."
passage_embedding = model.encode(passage)

# Calculate similarity
from sklearn.metrics.pairwise import cosine_similarity
similarity = cosine_similarity([query_embedding], [passage_embedding])[0][0]
print(f"Similarity: {similarity:.4f}")

For RAG Applications

from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer('YOUR_USERNAME/kazembed-v5')

# Your document corpus
documents = [
    "Астана — Қазақстан Республикасының астанасы.",
    "Алматы — Қазақстанның ең үлкен қаласы.",
    "Қазақстан — Орталық Азиядағы мемлекет.",
]

# Encode documents (do once, store in vector DB)
doc_embeddings = model.encode(["passage: " + doc for doc in documents])

# Query
query = "Қазақстанның астанасы қай қала?"
query_embedding = model.encode("query: " + query)

# Find most similar
similarities = np.dot(doc_embeddings, query_embedding)
best_idx = np.argmax(similarities)
print(f"Best match: {documents[best_idx]}")

Training Details

Training Data

Training Configuration

• Base Model: intfloat/multilingual-e5-base
• Epochs: 2
• Batch Size: 16
• Learning Rate: 1e-5
• Loss: MultipleNegativesRankingLoss
• Hardware: NVIDIA GPU

Training Strategy

We found that:

1. Retrieval-only data works best (no NLI/STS data)
2. 2 epochs is optimal (1 = underfit, 3 = overfit)
3. Larger batch size (16) provides more in-batch negatives

Limitations

• Optimized for Kazakh; performance on other languages may vary
• Best for retrieval tasks; may not be optimal for semantic similarity
• Requires query: and passage: prefixes for best results

Citation

@misc{kazembed2024,
  title={KazEmbed-V5: A Fine-tuned Embedding Model for Kazakh Language Retrieval},
  author={Your Name},
  year={2024},
  howpublished={HuggingFace Hub}
}

Acknowledgments

• Base model: intfloat/multilingual-e5-base
• Training data: ISSAI for KazQAD dataset