| Property | Value |
|---|---|
| Model Size | 14B parameters |
| Author | sometimesanotion |
| Model URL | Hugging Face |
| Architecture Type | Fusion Merge Model |
What is Lamarck-14B-v0.7-Fusion?
Lamarck-14B-v0.7-Fusion is an experimental language model that implements a novel multi-stage merge approach using the arcee_fusion method. The model combines various iterations of Lamarck and Qwenvergence models through a sophisticated four-step merge process, with three steps utilizing fusion techniques.
Implementation Details
The model architecture involves a complex merger of multiple models including Lamarck-14B-v0.7, Lamarckvergence, and Qwenvergence-14B-v12-Prose-DS. It uses bfloat16 precision and implements specific parameter configurations including int8_mask and normalization techniques.
- Multi-stage fusion process with SLERP merging
- Integration with Chocolatine-2-14B-Instruct-v2.0.3
- Enhanced emphasis on Qwenvergence-14B-v12-Prose-DS in later layers
- Specialized layer configurations for attention and MLP components
Core Capabilities
- Strong GPQA (General Purpose Question Answering) performance
- Enhanced reasoning capabilities
- Superior prose generation
- Optimized for free-form creative tasks
- Moderate IFEVAL performance
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its innovative use of the arcee_fusion merge method in a multi-stage process, combining the strengths of multiple parent models while maintaining high-quality prose generation capabilities.
Q: What are the recommended use cases?
This model is particularly well-suited for free-form creative tasks, general-purpose question answering, and applications requiring strong reasoning capabilities. It excels in scenarios where high-quality prose generation is essential.