Open Korean LLM - KorNAT-Harmless: leaderboard
Ko-Harmlessness (KorNAT family, SELECTSTAR and KAIST AI): multiple-choice harmlessness items, scored by normalized accuracy on a 0-100 scale; higher is better. A perfect score with zero standard error is dropped as a harness artifact. From Upstage and NIA's Open Ko-LLM Leaderboard 2 (lm-evaluation-harness on private Korean test sets); frozen since its last evaluations in January 2025.
Metric: Normalized accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 1157 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Myrrh_solar_10.7b_3.0 | 88.8 | #1055 |
| 2 | free-solar-evo-v0.13 | 85.08 | #1077 |
| 3 | free-solar-evo-v0.11 | 83.85 | #1144 |
| 4 | free-solar-evo-v0.1 | 83.83 | #1152 |
| 5 | NeuralMonarch-7B | 79.7 | #944 |
| 6 | AlphaMonarch-7B | 79.63 | #1042 |
| 7 | nox-solar-10.7B-v4 | 78.26 | #1250 |
| 8 | multi_verse_model | 77.98 | #1097 |
| 9 | MixTAO-7Bx2-MoE-v8.1 | 77.91 | #812 |
| 10 | lambda-qwen2.5-14B-dpo-test | 77.32 | #481 |
| 11 | NeuralSynthesis-7B-v0.1 | 76.74 | #973 |
| 12 | Beyonder-4x7B-v3 | 76.71 | #943 |
| 13 | dzakwan-MoE-4x7B-Beta | 76.71 | #963 |
| 14 | BeagSake-7B | 75.05 | #984 |
| 15 | CarbonBeagle-11B-truthy | 74.41 | #1023 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=open-korean-llm-kornat-harmless · How It Works · Data refreshed daily, snapshot 2026-11-05.