Darwin-27B-KR-V3

K-AI Leaderboard optimized model with CLIcK-heavy SFT data.

Training Details

• Base Model: FINAL-Bench/Darwin-27B-Opus
• Method: LoRA SFT (r=64, alpha=128)
• Data: 1,500 examples (CLIcK-heavy: 600 CLIcK + 300 KoMMLU-Pro + 300 MuSR-Ko + 300 Com2-main-Ko)
• Epochs: 3
• Architecture: Qwen3.5-27B

Key Features

• Enhanced CLIcK (Korean Cultural Literacy) performance through double CLIcK training data
• Strong KMMLU-Pro performance maintained
• Reasoning-capable with thinking mode support