KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024 Overview

KISTI-KONI/KONI-Llama3.1-8B-Instruct-20241024 is an 8 billion parameter instruction-tuned large language model developed by the Korea Institute of Science and Technology Information (KISTI). This model is a specialized variant of the Llama 3.1 family, specifically engineered for science and technology applications.

Key Capabilities & Features

• Domain Specialization: Explicitly trained on a vast corpus of scientific and technological data, making it highly proficient in these fields.
• Enhanced Performance: Demonstrates significantly improved performance compared to its earlier iterations.
• Base Model: Built upon a merged foundation of Meta-Llama-3-8B and KISTI-KONI/KONI-Llama3.1-8B-20240824.
• Alignment: Utilizes both Supervised Fine-Tuning (SFT) with approximately 11k data points and Direct Preference Optimization (DPO) with 7k data points for robust instruction following.
• Multilingual Data: SFT and DPO datasets include internally generated data, publicly available data, and translated/curated data (e.g., from argilla/dpo-mix-7k), with Korean translations where necessary.

Benchmark Performance

Evaluated on the LogicKor benchmark, the model achieved an Overall score of 8.93, with strong performance across various categories:

• Reasoning: 8.15
• Math: 8.79
• Writing: 9.22
• Coding: 9.21
• Comprehension: 9.65
• Grammar: 8.57
• Single-turn: 9.05
• Multi-turn: 8.81

Ideal Use Cases

This model is particularly well-suited for developers and researchers working on applications requiring deep understanding and generation within scientific, engineering, and technological contexts, especially those involving Korean language data. Its specialized training makes it a strong candidate for tasks like technical documentation, scientific inquiry, code assistance, and complex problem-solving in these domains.