Sarashina2.2-3B-Instruct-v0.1
| Property | Value |
|---|---|
| Model Type | Autoregressive Language Model |
| Parameter Count | 3 Billion |
| Primary Language | Japanese |
| License | MIT |
| Author | SB Intuitions |
What is sarashina2.2-3b-instruct-v0.1?
Sarashina2.2-3B-Instruct is an advanced Japanese language model developed by SB Intuitions, specifically designed for instruction-following tasks. The model demonstrates exceptional performance in both Japanese and English language benchmarks, surpassing many comparable models in its class.
Implementation Details
The model is built on an autoregressive architecture and is optimized for both Japanese and cross-lingual tasks. It can be easily implemented using the Transformers library with bfloat16 precision and supports automatic device mapping for efficient resource utilization.
- Achieves 6.51 score on Japanese MT Bench
- Scores 7.71 on English MT Bench
- Outperforms other models in its size range including Qwen2.5-3B and llm-jp-3-3.7b
Core Capabilities
- High-quality Japanese language understanding and generation
- Strong cross-lingual performance in English tasks
- Instruction-following with contextual awareness
- Flexible deployment options with HuggingFace integration
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its exceptional performance in both Japanese and English tasks, achieving some of the highest scores in the ELYZA-tasks-100 and MT Bench evaluations among models of similar size. It's particularly notable for maintaining high performance across language boundaries.
Q: What are the recommended use cases?
The model is well-suited for Japanese language processing tasks, including text generation, instruction following, and cross-lingual applications. However, users should note that it has limited safety training and may require additional fine-tuning for specific use cases.
