NQLSG-Qwen2.5-14B-OriginalFusion
| Property | Value |
|---|---|
| Model Size | 14B parameters |
| Base Model | Qwen2.5-14B |
| Precision | bfloat16 |
| HuggingFace | Link |
What is NQLSG-Qwen2.5-14B-OriginalFusion?
NQLSG-Qwen2.5-14B-OriginalFusion is an advanced language model created through a sophisticated merger of multiple high-performing models, primarily based on Qwen2.5-14B architecture. This model represents a careful fusion of nine distinct models, including various versions of Qwen2.5-14B, Lamarck variants, and DeepSeek, using the Model Stock merge method.
Implementation Details
The model utilizes a unique merging strategy with bfloat16 precision and implements int8_mask for parameter optimization. It's built upon the NQLSG-Qwen2.5-14B-MegaFusion-v8 as its foundation, incorporating both instruction-tuned and specialized coding variants of the base models.
- Implements Model Stock merge methodology
- Utilizes bfloat16 precision for optimal performance
- Incorporates automatic chat template integration
- Features int8 masking for parameter efficiency
Core Capabilities
- Enhanced instruction following through multiple instruction-tuned model integration
- Specialized coding capabilities from Qwen2.5-Coder-14B
- Improved reasoning from Lamarck model variants
- Balanced performance through strategic model fusion
Frequently Asked Questions
Q: What makes this model unique?
This model's uniqueness lies in its comprehensive fusion of nine carefully selected models, combining the strengths of Qwen2.5, Lamarck, and DeepSeek architectures while maintaining efficient parameter usage through int8 masking.
Q: What are the recommended use cases?
The model is well-suited for general-purpose tasks, coding applications, and instruction-following scenarios, thanks to its diverse model heritage and specialized component models.
