CarrotAI/Llama-3.2-Rabbit-Ko-3B-Instruct
CarrotAI/Llama-3.2-Rabbit-Ko-3B-Instruct is a 3.2 billion parameter instruction-tuned large language model developed by CarrotAI, based on the Llama-3.2 architecture. It is fine-tuned using high-quality Korean and English datasets, making it proficient in both languages. This model is designed for general instruction-following tasks, offering a compact solution for applications requiring bilingual capabilities. It demonstrates performance on benchmarks like gsm8k, ifeval, haerae, and kobest, particularly excelling in Korean language understanding and generation.
Loading preview...
Carrot Llama-3.2 Rabbit Ko 3B Instruct
CarrotAI/Llama-3.2-Rabbit-Ko-3B-Instruct is a 3.2 billion parameter instruction-tuned large language model developed by CarrotAI. This model is built upon the Llama-3.2 architecture and has undergone Supervised Fine-Tuning (SFT) using high-quality datasets in both Korean and English, making it a strong candidate for bilingual applications.
Key Capabilities
- Bilingual Proficiency: Fine-tuned with extensive Korean and English data, enabling effective communication and understanding in both languages.
- Instruction Following: Designed to accurately follow instructions for a variety of tasks.
- Compact Size: At 3.2 billion parameters, it offers a balance between performance and computational efficiency, suitable for deployment in resource-constrained environments.
Performance Highlights
The model's performance is evaluated across several benchmarks:
- gsm8k: Achieves 0.6490 exact match on English gsm8k and 0.3275 on Korean gsm8k.
- ifeval: Demonstrates an instruction-level loose accuracy of 0.8058.
- haerae: Shows an overall accuracy of 0.4180 on Korean general knowledge tasks.
- kobest: Scores 0.7664 accuracy on kobest_boolq and 0.5620 on kobest_copa, indicating solid performance in Korean natural language understanding tasks.
Limitations
Due to its 3B parameter scale, the model may exhibit limited performance on highly complex tasks and may lack deep expertise in specific domains. Users should also be aware of potential biases and hallucination risks inherent in LLMs.