allganize/Llama-3-Alpha-Ko-8B-Instruct

Alpha-Instruct: A Bilingual Llama-3 Model for Korean and English

Alpha-Instruct, developed by Allganize, is an 8 billion parameter language model built on the Llama-3 architecture. It leverages an innovative Evolutionary Model Merging technique, combining Meta-Llama-3-8B, Meta-Llama-3-8B-Instruct, and Llama-3-Open-Ko-8B to achieve strong bilingual performance.

Key Capabilities & Differentiators

• Exceptional Logical Reasoning: Alpha-Instruct scores 6.62 on the LogicKor benchmark, a performance comparable to 70B models, highlighting its advanced computational and reasoning skills.
• Bilingual Proficiency: The model demonstrates strong capabilities in both Korean and English, making it suitable for diverse language tasks.
• Human Preference Optimization: It was refined using high-quality, curated datasets like Korean-Human-Judgements and Orca-Math, and optimized with ORPO for improved human preference scores and real-life applicability.
• Community-Driven Development: Allganize emphasizes a community-based approach, drawing inspiration from various communities and committing to sharing insights on data, methods, and models.

Benchmark Performance

Alpha-Instruct shows competitive results on Korean benchmarks:

• LogicKor: Achieved an overall score of 6.62, outperforming other 8B models like MLP-KTLim/llama-3-Korean-Bllossom-8B and Alpha-Ko-Evo.
• KoBEST: Demonstrates strong performance across various tasks, including kobest_boolq (0.8369) and kobest_sentineg (0.9244), indicating robust understanding and generation capabilities in Korean.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

• Complex Korean and English language understanding and generation.
• Tasks demanding strong logical reasoning and problem-solving.
• Scenarios where efficiency (8B parameters) is crucial without sacrificing high-quality output.