Noid3a-Labs/Sparky-4B-V1
Sparky-4B-V1 by Noid3a Labs is an ultra-efficient, high-discipline 3.1 billion parameter instruction model based on the Qwen2.5-3B architecture, featuring a 32K context length. It is fine-tuned for strict structural discipline, zero-waffle response generation, and rapid inference. This model excels in code generation, agentic tool workflows, and mathematical reasoning, outperforming its base model in Python code generation and instruction following.
Loading preview...
Sparky-4B-V1: An Efficient and Disciplined Instruction Model
Sparky-4B-V1 is a 3.1 billion parameter instruction-tuned model developed by Noid3a Labs, built upon the Qwen2.5-3B base architecture. It is specifically designed for high discipline, zero-waffle response generation, and rapid inference execution, making it suitable for tasks where precision and efficiency are paramount.
Key Capabilities & Performance
- Enhanced Code Generation: Sparky-4B-V1 demonstrates strong performance in Python code generation, achieving 51.22% on HumanEval (0-Shot,
pass@1), significantly outperforming the Qwen 2.5 3B base model's 42.10%. - Improved Instruction Following: The model excels in instruction adherence, scoring 47.32% on IFEval (0-Shot, Strict), compared to Qwen 2.5 3B Instruct's 42.50%.
- Agentic Workflows: Its disciplined response generation makes it well-suited for agentic tool use and structured tasks.
- Mathematical Reasoning: While showing strong overall performance, its GSM8K score for multi-step math reasoning is 61.03% (5-Shot, Flexible Extract).
Use Cases
Sparky-4B-V1 is ideal for applications requiring:
- Code generation and completion.
- Strict instruction following and assistant tasks.
- Agentic tool integration where precise and structured outputs are critical.
Technical Details
The model utilizes the standard ChatML prompt format for interaction. It can be run locally using Ollama, with specific instructions provided in the original README.