🦉 CodeSearch-ModernBERT-Owl-Plus: High-Performance Sentence-BERT for Code Search
CodeSearch-ModernBERT-Owl-Plus is a high-performance code search model fine-tuned in a Sentence-BERT architecture, based on the pretrained CodeModernBERT-Owl v1.0.
This model is optimized for function-level search within codebases and natural language queries, achieving state-of-the-art results on the MTEB benchmark.
🛠 Features
- ✅ Fine-tuned in Sentence-BERT format from CodeModernBERT-Owl
- ✅ Supports multiple languages (Python, Java, JavaScript, etc.)
- ✅ Specialized encoder for high-accuracy code search
- ✅ Ideal for multi-stage (dual encoder) retrieval setups
- ✅ Generates rich semantic embeddings for code and queries
📊 Evaluation on MTEB Benchmark
🏆 Main Scores in MTEB
This model achieved the following main scores (based on NDCG@10):
- CodeSearchNetRetrieval:
main_score = 0.8918 - COIR-CodeSearchNetRetrieval:
main_score = 0.8013
🧪 CodeSearchNetRetrieval (MTEB)
| Metric | Score |
|---|---|
| MRR@10 | 0.8704 |
| NDCG@10 | 0.8918 |
| MAP@10 | 0.8704 |
| Recall@10 | 0.9563 |
| Precision@10 | 0.0956 |
This model achieves strong performance across all ranking metrics and demonstrates balanced retrieval capability.
🧪 COIR-CodeSearchNetRetrieval (MTEB)
| Metric | Score |
|---|---|
| MRR@10 | 0.7751 |
| NDCG@10 | 0.8013 |
| MAP@10 | 0.7751 |
| Recall@10 | 0.8826 |
| Precision@10 | 0.0883 |
Robust and consistent performance is also maintained on the COIR dataset, demonstrating strong generalization.
📥 Usage Example
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("Shuu12121/CodeSearch-ModernBERT-Owl-Plus")
embeddings = model.encode(["binary search function", "def binary_search(arr, target): ..."])
📝 Conclusion
- ✅ An optimized Sentence-BERT model based on CodeModernBERT-Owl
- ✅ Achieves MRR@10 > 0.87 on MTEB CodeSearchNetRetrieval
- ✅ Ready for integration in production-level code search systems
📜 License
📄 Apache-2.0
📧 Contact
For questions or inquiries, feel free to reach out: 📧 shun0212114@outlook.jp