Yi-34B-200K-RPMerge
| Property | Value |
|---|---|
| Parameter Count | 34.4B |
| License | Yi License |
| Tensor Type | BF16 |
| Context Length | 200K |
| Merge Method | DARE TIES |
| Base Papers | DARE, TIES |
What is Yi-34B-200K-RPMerge?
Yi-34B-200K-RPMerge is a sophisticated model merge specifically designed for enhanced storytelling and long-context processing. It combines several carefully selected Yi-34B models, primarily using Vicuna format, to create a cohesive system capable of handling context lengths up to 40K+ tokens while maintaining high-quality narrative generation.
Implementation Details
The model utilizes the DARE TIES merge method, incorporating six major models including Tess-34B, Nous-Capybara-34B, and ChatAllInOne. Each component is weighted specifically to maintain optimal performance, with weights ranging from 0.05 to 0.19 and varying density parameters.
- Implements Vicuna prompt format for consistency
- Supports context lengths of 40K-90K on 24GB GPUs using exllamav2
- Utilizes quadratic sampling for improved output quality
- Features carefully tuned parameters to handle Yi's large tokenizer vocabulary
Core Capabilities
- Extended context processing up to 200K tokens
- Enhanced storytelling and narrative generation
- Robust instruction-following abilities
- Multi-character story support
- Reduced likelihood of generating refusals
- Efficient performance on consumer hardware with proper optimization
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its focused approach to storytelling while maintaining extended context capabilities. Unlike kitchen sink merges, it specifically uses Vicuna-format models to ensure consistency and avoid format conflicts.
Q: What are the recommended use cases?
The model excels at long-form narrative generation, creative writing, and extended context analysis. It's particularly well-suited for novel continuations, multi-character storytelling, and general instruction-following tasks requiring extended context understanding.
