Hugging Face Overview
Overview
HaiJava-Surgeon-Qwen2.5-Coder-7B-SFT-v1 is a 7.6 billion parameter model, fine-tuned from the Qwen/Qwen2.5-Coder-7B-Instruct base model. It has undergone 1000 training steps specifically on Java bug-fixing samples, resulting in a merged model that offers enhanced performance for Java code correction tasks. This model is ready for production deployment and serves as a stronger starting point for subsequent fine-tuning efforts.
Key Capabilities
- Improved Java Bug Fixing: Achieves an 82.28% pass@1 on the MultiPL-E Java benchmark, a notable +14.56% improvement over the base model, solving 23 more problems. Internal evaluations show a +55.6% increase in overall accuracy on a 50-sample test set.
- Enhanced Syntax and Logic Correction: Demonstrates significant improvements in fixing syntax errors (+50%) and logic bugs (+33%) in Java code.
- Direct Inference: The model is a merged LoRA + base model, allowing for faster loading and simpler deployment without the need for separate adapter files.
- Strong Baseline for Further Training: Can be used as an improved base model for continued fine-tuning with new LoRA adapters, starting from a higher accuracy baseline (28% vs 18%).
Use Cases
- Production Deployment: Ideal for direct integration into applications requiring robust Java bug-fixing capabilities.
- Continued Fine-tuning: Suitable for developers looking to specialize the model further for specific Java domains or new bug types.
- Code Quality Tools: Can be integrated into IDEs or CI/CD pipelines for automated code review and bug detection in Java projects.
Limitations
- The model still struggles with API misuse detection, edge case handling, and Null Pointer Exception fixes, indicating areas for future training focus.
- It shows no proficiency in Python bug fixing, as it is specialized for Java.