Zamba2-2.7B-instruct
| Property | Value |
|---|---|
| Parameter Count | 2.7B |
| License | Apache 2.0 |
| Model Type | Hybrid SSM-Transformer |
| Tensor Type | F32/BF16 |
What is Zamba2-2.7B-instruct?
Zamba2-2.7B-instruct is a groundbreaking hybrid model that combines state-space modeling (Mamba2) with transformer architecture. Fine-tuned on multiple instruction-following and chat datasets, it demonstrates exceptional performance that surpasses many larger models, including Mistral-7B-Instruct and Gemma2-2B-Instruct.
Implementation Details
The model architecture features a unique backbone of Mamba2 layers interleaved with shared attention layers. It implements LoRA projection matrices for the shared MLP, enabling position-specific specialization while maintaining minimal parameter overhead. The model has been fine-tuned through a two-step process: initial SFT training on ultrachat_200k and Infinity-Instruct, followed by DPO training on multiple preference datasets.
- Innovative hybrid architecture combining SSM and transformer blocks
- Efficient parameter sharing through shared attention mechanisms
- Enhanced information flow through embedding concatenation
- Optimized for both performance and efficiency
Core Capabilities
- Superior instruction-following abilities (MT-Bench score: 72.40)
- Extremely low inference latency
- Reduced memory footprint compared to traditional transformers
- Excellent performance in reasoning tasks
- Efficient on-device deployment capabilities
Frequently Asked Questions
Q: What makes this model unique?
The model's hybrid architecture combining Mamba2 state-space modeling with transformer blocks enables exceptional performance while maintaining low computational requirements. It achieves better results than many larger models while using fewer parameters.
Q: What are the recommended use cases?
The model is particularly well-suited for on-device applications requiring strong instruction-following capabilities, rapid response times, and efficient resource usage. It excels in general-purpose text generation, reasoning tasks, and conversational applications.
