jiwoochris/ko-llama2-v1

Model Overview

jiwoochris/ko-llama2-v1 is a 13 billion parameter Korean-English Llama 2 model developed by Jiwoo Chris Jung. It was instruction-tuned using LoRA, building upon the hyunseoki/ko-en-llama2-13b base model, which itself is a Llama 2 13B model with continued Korean-English pretraining. The model was released in October 2023.

Key Differentiators & Performance

This model's primary distinction lies in its instruction-tuning methodology, which prioritized data quality over volume. It was fine-tuned on a carefully selected set of 1,944 instruction examples, following the principles of "less-is-more" as observed in research like LIMA and Instruction Mining. This approach proved highly effective, as the model achieved the #1 ranking on the Open Ko-LLM Leaderboard on October 28-29, 2023, outperforming 230 other models in Korean LLM benchmarks.

Intended Use Cases

This model is particularly well-suited for applications requiring high-quality instruction-following in both Korean and English. Its strong performance on the Open Ko-LLM Leaderboard suggests its efficacy in understanding and generating relevant responses for Korean-centric tasks. Developers can leverage its capabilities for various NLP applications where robust bilingual understanding and generation are crucial.

Limitations

As a research checkpoint from October 2023, ko-llama2-v1 was not specifically trained for safety. It may produce incorrect or fabricated content and inherits the limitations and license terms of its Llama 2 base model.