| Property | Value |
|---|---|
| Parameter Count | 69B |
| License | CC-BY-NC-2.0 |
| Model Type | Text Generation |
| Architecture | LLaMA2-based merge |
| Tensor Type | F32/BF16 |
What is lzlv_70b_fp16_hf?
lzlv_70b_fp16_hf is a sophisticated multi-model merge combining three powerful 70B parameter language models: Nous-Hermes-Llama2-70b, Xwin-LM-7B-V0.1, and Doctor-Shotgun/Mythospice-70b. The model aims to balance creative capabilities with robust instruction-following, making it particularly suitable for roleplay and creative tasks.
Implementation Details
The model employs a complex merging strategy using SLERP gradients. The process involves two primary components: first merging Mythospice with Xwin using gradients [0.25, 0.3, 0.5], then merging Xwin with Hermes using gradients [0.4, 0.3, 0.25]. These components are finally combined using a 0.5 SLERP weight.
- Uses Vicuna prompt format
- Supports both F32 and BF16 tensor types
- Quantized versions available through TheBloke's GGUF conversions
Core Capabilities
- Enhanced creative text generation
- Strong instruction-following abilities inherited from Xwin-LM
- Balanced performance in complex scenarios
- Improved narrative and roleplay capabilities
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its careful balance of creative capabilities and instruction-following abilities, achieved through a sophisticated merging process of three distinct 70B models.
Q: What are the recommended use cases?
The model excels in creative writing, roleplaying scenarios, and complex narrative tasks while maintaining the ability to follow detailed instructions effectively.