Model Overview

Synatra-7B-Instruct-v0.2 is a 7 billion parameter instruction-tuned language model developed by StableFluffy, building upon the mistralai/Mistral-7B-Instruct-v0.1 base model. It is designed to improve performance on Korean language tasks, featuring an 8192-token context length.

Key Capabilities & Performance

This model shows enhanced capabilities in Korean language understanding and reasoning, as evidenced by its benchmark results on the Ko-LLM-Leaderboard:

• Ko-MMLU: Achieves 43.99%
• Ko-TruthfulQA: Achieves 45.77%
• Overall Average: 44.78%

Compared to its predecessor, Synatra-V0.1-7B-Instruct, this version demonstrates improvements across most Korean benchmarks, particularly in MMLU and TruthfulQA, though it shows relative weakness in Ko-CommonGen V2.

Usage & Licensing

The model is intended for non-commercial use only under the CC BY-NC 4.0 license. Commercial use requires direct contact with the developer. It follows a specific instruction format, requiring prompts to be enclosed within [INST] and [/INST] tokens for optimal performance with instruction fine-tuning.