DireDreadlord/Dragon-1-0.5B
DireDreadlord/Dragon-1-0.5B is a 0.5 billion parameter code reasoning and generation model built upon the Qwen2-0.5B-Instruct architecture. It is specifically SFT and RL trained on code reasoning traces to provide accurate and hallucination-free code snippets and long-form code generation across major programming languages. This lightweight model excels at instruction following and advanced reasoning for code-related subjects, making it suitable for deployment on commercial-grade GPUs.
Loading preview...
Overview
DireDreadlord/Dragon-1-0.5B is a lightweight 0.5 billion parameter model from the Dragon-1 series, designed for enhanced code reasoning and generation. Built on the Qwen2 architecture, it specializes in producing accurate and hallucination-free code snippets and long-form code in various programming languages. Its small size allows for efficient execution on common laptop and commercial GPUs.
Key Capabilities
- Code Reasoning: Enhanced reasoning capabilities for complex coding problems, trained on the deepseek-v4-reasoning-code-2500 dataset.
- Code Generation: Accurate and quick generation of both short code snippets and extensive code blocks from natural language instructions.
- Instruction Following: Proficient in understanding and executing code-related instructions.
- Subject Matter Expertise: Acts as a Q/A and subject matter expert for code-related topics.
- Efficient Deployment: Optimized for performance on commercial-grade GPUs due to its compact size.
Training Details
The model underwent a two-phase training process. Phase 1 involved 5,000 steps of Supervised Fine-Tuning (SFT) on the deepseek-v4-reasoning-code-2500 dataset. Phase 2 utilized a GRPO algorithm for 500 steps of Reinforcement Learning (RL) on the same dataset, further refining its reasoning and generation quality.