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

allganize/Llama-3-Alpha-Ko-8B-Evo: An Evolutionary Bilingual Base Model

Alpha-Ko-8B-Evo is an 8 billion parameter language model developed by allganize, leveraging the Evolutionary Model Merging technique. This model serves as a foundational 'Evo' version, intended for further fine-tuning to specific tasks, while its instruction-tuned counterpart, Alpha-Instruct, is recommended for general chat.

Key Capabilities & Development:

• Bilingual Proficiency: Demonstrates strong capabilities in both Korean and English.
• Evolutionary Model Merging: Developed by merging Meta-Llama-3-8B, Meta-Llama-3-8B-Instruct, and Llama-3-Open-Ko-8B, using a 1:1 ratio of task-specific datasets from KoBEST and Haerae.
• Enhanced Human Preference: Utilizes specialized datasets, including Korean-Human-Judgements and Orca-Math, to 'heal' model output and boost human preference scores.
• Logical Reasoning: Achieved an impressive 6.60 on the LogicKor benchmark, rivaling 70B models in performance, indicating advanced computational and reasoning skills.

Benchmark Highlights:

• LogicKor: Alpha-Ko-Evo scored 5.190 overall, showcasing strong logical reasoning. The instruction-tuned variant, Alpha-Ko-Instruct, achieved an even higher 6.600.
• KoBEST (5-shot accuracy): Demonstrated competitive performance, with an overall score of 0.7229, notably excelling in kobest_boolq (0.8547) and kobest_sentineg (0.9471).

This model is ideal for developers looking for a robust bilingual base model that combines strong reasoning with human-aligned output, particularly for applications requiring high performance in Korean and English language tasks.