Open Korean LLM - Ko-GSM8K (flexible extract): leaderboard
Korean GSM8K grade-school math word problems (Flitto), scored by exact match of the final number using lm-evaluation-harness's flexible extraction; the Space's own column uses strict matching, which scores correct free-form answers as zero. 0-100, higher is better. 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: Exact match, flexible extract (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 1158 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gemma-2-27B-it-SimPO-37K-100steps | 78.24 | #609 |
| 2 | lambda-qwen2.5-14B-dpo-test | 77.71 | #481 |
| 3 | Gemma 2 27B (IT) | 77.63 | #652 |
| 4 | gemma2-gutenberg-27B | 77.18 | #453 |
| 5 | Qwen2.5-14B-Gutenberg-Instruct-Slerpeno | 76.35 | #431 |
| 6 | Qwen2.5-14B-Gutenberg-1e-Delta | 76.19 | #473 |
| 7 | Qwen 2.5 14B Instruct | 75.97 | #654 |
| 8 | Qwen2.5-Lumen-14B | 75.82 | #482 |
| 9 | Rombos-LLM-V2.6-Qwen-14B | 75.59 | #520 |
| 10 | qwen2.5-reinstruct-alternate-lumen-14B | 75.36 | #445 |
| 11 | SuperNova-Medius | 75.21 | #483 |
| 12 | T.E-8.1 | 72.71 | #695 |
| 13 | Qwen 2.5 Coder 14B Instruct | 72.18 | #685 |
| 14 | gemma-2-9B-it-SimPO | 72.02 | #569 |
| 15 | Qwen2.5-7B-Instruct-Uncensored | 71.34 | #684 |
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-ko-gsm8k-flexible-extract · How It Works · Data refreshed daily, snapshot 2026-11-05.