Open Korean LLM - KorNAT-Helpful: leaderboard
Ko-Helpfulness (KorNAT family, SELECTSTAR and KAIST AI): multiple-choice helpfulness 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 | 82.13 | #1055 |
| 2 | free-solar-evo-v0.1 | 77.53 | #1152 |
| 3 | free-solar-evo-v0.13 | 76.8 | #1077 |
| 4 | free-solar-evo-v0.11 | 75.38 | #1144 |
| 5 | nox-solar-10.7B-v4 | 67.5 | #1250 |
| 6 | lambda-qwen2.5-14B-dpo-test | 66.4 | #481 |
| 7 | Qwen2.5-Lumen-14B | 60.9 | #482 |
| 8 | Qwen 2.5 14B Instruct | 60.65 | #654 |
| 9 | qwen2.5-reinstruct-alternate-lumen-14B | 60.48 | #445 |
| 10 | Qwen2.5-14B-Gutenberg-1e-Delta | 60.2 | #473 |
| 11 | Qwen2.5-14B-Gutenberg-Instruct-Slerpeno | 60.05 | #431 |
| 12 | Rombos-LLM-V2.6-Qwen-14B | 59.2 | #520 |
| 13 | SuperNova-Medius | 58.63 | #483 |
| 14 | Qwen2.5-7B-Instruct-abliterated | 56.28 | #792 |
| 15 | Llama-3-Base-8B-SFT-RDPO | 56.23 | #972 |
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-helpful · How It Works · Data refreshed daily, snapshot 2026-11-05.