DireDreadlord/Dragon-1.5-0.5B

Dragon-1.5-0.5B Overview

DireDreadlord/Dragon-1.5-0.5B is a lightweight, 0.5 billion parameter language model built upon the Qwen2 architecture, specifically the Qwen2-0.5B-Instruct base. It is designed to provide accurate and quick text generation across various topics, including code-related problems. This model distinguishes itself with enhanced reasoning capabilities, making it suitable for complex instruction following and problem-solving.

Key Capabilities

• Advanced Reasoning: Incorporates high-end reasoning capabilities, trained on specialized reasoning traces.
• Instruction Following: Excels at understanding and executing user instructions.
• Code Generation: Offers enhanced code snippet and long-form code generation.
• Q/A and SME: Provides subject matter expert capabilities on general and code-related subjects.
• Efficient Performance: Its small size allows comfortable operation on most laptop/commercial-grade GPUs.

Training Details

The model underwent a two-phase training process:

• Phase 1 (SFT): Supervised fine-tuning on the deepseek-v4-distill-8000x dataset (approximately 7,716 rows) for 4,000 steps.
• Phase 2 (RL): Reinforcement learning via a GRPO algorithm on the deepseek-v4-reasoning-code-2500 dataset (approximately 7,716 rows) for 350 steps. This phase specifically endows the model with enhanced reasoning and aims for higher quality, hallucination-free generations.

Good For

• Applications requiring advanced reasoning in a compact model.
• Scenarios needing efficient code generation and problem-solving.
• Use cases where low computational footprint is critical, such as deployment on commercial GPUs.