KISTI-KONI/KONI-Llama3.1-70B-Instruct-20241115: Science & Technology Specialized LLM

KONI (KISTI Open Natural Intelligence) is a large language model developed by the Korea Institute of Science and Technology Information (KISTI), specifically designed for science and technology domains. This 70 billion parameter instruction-tuned model is built on the Meta-Llama-3.1-70B-Instruct base.

Key Capabilities & Features

• Science and Technology Specialization: Explicitly trained on a vast, specialized corpus of scientific and technological data, making it highly effective for domain-specific tasks.
• Enhanced Performance: Demonstrates significant performance improvements over previous KONI versions.
• Fine-tuning: Utilizes Supervised Fine-Tuning (SFT) with approximately 11k data points and Direct Preference Optimization (DPO) with 7k data points, including internally generated and publicly available data, with Korean translations.

Benchmark Performance

Evaluated using the LogicKor benchmark, KONI-Llama3.1-70B-Instruct-20241115 achieved an Overall Score of 9.38, with strong results in:

• Reasoning: 9.07
• Math: 9.65
• Coding: 9.65
• Comprehension: 9.86

Ideal Use Cases

This model is particularly well-suited for applications requiring deep understanding and generation within scientific and technological contexts, including research assistance, technical documentation, and specialized Q&A in these fields.