| Property | Value |
|---|---|
| Author | Sao10K |
| License | cc-by-nc-4.0 |
| Model Size | 12B parameters |
| Base Architecture | Mistral-NeMo |
| HuggingFace | Link |
What is MN-12B-Lyra-v4?
MN-12B-Lyra-v4 is an advanced language model that builds upon previous Lyra versions, specifically designed to enhance instruction following and coherency. It represents a significant evolution in the Lyra series, implementing a unique reinforcement learning approach that targets instruction handling directly on the base NeMo model rather than using traditional SFT-first methodology.
Implementation Details
The model features comprehensive ChatML support and its variants, with specific attention to tokenizer optimization and quantization improvements. It implements a sophisticated sampling strategy with recommended temperature ranges of 0.6-1.0 and crucial min_p values of 0.1-0.2 for optimal NeMo performance.
- Enhanced tokenizer configuration with improved stability
- Support for multiple chat template formats
- Optimized stopping string handling
- Fixed token generation issues while maintaining core functionality
Core Capabilities
- Advanced instruction following and coherency
- Flexible chat template support including ChatML and its variants
- Improved quantization handling
- Robust response generation with optimized sampling parameters
Frequently Asked Questions
Q: What makes this model unique?
The model's unique approach lies in its direct reinforcement learning implementation targeting instruction and coherency on the base NeMo model, rather than using traditional SFT-first approaches. This has resulted in improved quantization handling and more stable performance.
Q: What are the recommended use cases?
The model is particularly well-suited for applications requiring structured dialogue interactions, thanks to its comprehensive ChatML support and enhanced instruction-following capabilities. It's optimized for scenarios requiring coherent and context-aware responses.