| Property | Value |
|---|---|
| Parameter Count | 22.1B |
| License | MIT |
| Tensor Type | BF16 |
| Context Length | 4K tokens |
| Language | English |
What is solar-pro-preview-instruct?
Solar Pro Preview is an advanced large language model that represents a significant breakthrough in efficient AI deployment. Developed by Upstage, this 22B parameter model is specifically designed to run on a single GPU with 80GB VRAM, while delivering performance that rivals models three times its size. It's built using an enhanced depth up-scaling method, transforming a Phi-3-medium model from 14B to 22B parameters.
Implementation Details
The model utilizes the ChatML template for optimal performance in conversational and instruction-following tasks. It's implemented using the Transformers library and requires specific dependencies including torch 2.3.1 and flash_attn 2.5.8. The model particularly excels in benchmark performance, scoring 79.14 on MMLU and 84.37 on IFEval.
- Enhanced depth up-scaling architecture
- Optimized for single GPU deployment
- Carefully curated training strategy
- Supports instruction-tuned tasks
Core Capabilities
- Superior instruction-following abilities (84.37 on IFEval)
- Strong performance on mathematical reasoning (89.69 on GSM8K)
- Excellent knowledge assessment scores (79.14 on MMLU)
- Efficient resource utilization on single GPU
- Comprehensive benchmark performance across multiple domains
Frequently Asked Questions
Q: What makes this model unique?
This model uniquely combines efficiency with high performance, delivering capabilities comparable to 70B parameter models while requiring only a single GPU. Its enhanced depth up-scaling method and carefully curated training strategy enable superior performance in instruction-following and knowledge-based tasks.
Q: What are the recommended use cases?
The model is particularly well-suited for conversational AI applications, instruction-following tasks, and scenarios requiring strong reasoning capabilities. It's ideal for deployments where computational resources are limited but high performance is required.