Llama-3.2-Korean-GGACHI-1B-Instruct-v1 Overview

Llama-3.2-Korean-GGACHI-1B-Instruct-v1 is a 1 billion parameter instruction-tuned language model developed by torchtorchkimtorch. It is built upon the Llama-3.2-1B-Instruct architecture and has been specifically optimized for Korean language tasks. The model underwent fine-tuning using an extensive collection of over 230,000 high-quality Korean language datasets.

Key Capabilities and Performance

This model demonstrates enhanced performance on several Korean benchmarks (KOBEST) compared to its base Llama-3.2-1B-Instruct model. Notable improvements include:

• KOBEST COPA: Achieved 0.521 (0-shot), 0.549 (5-shot), and 0.549 (10-shot) accuracy, surpassing the base model.
• KOBEST HellaSwag: Showed improved accuracy with 0.380 (0-shot), 0.398 (5-shot), and 0.394 (10-shot).
• KOBEST Sentineg: Significantly outperformed the base model with 0.594 (0-shot), 0.795 (5-shot), and 0.821 (10-shot) accuracy.

While performance on kobest_boolq was comparable or slightly lower in some few-shot scenarios, the overall fine-tuning has yielded a model highly proficient in Korean-specific natural language understanding tasks.

Ideal Use Cases

This model is particularly well-suited for applications requiring robust performance in Korean language processing, such as:

• Korean text generation and summarization.
• Korean question answering and natural language inference.
• Developing AI agents or chatbots that interact primarily in Korean.