KULLM3: benchmark results
Korea University NLP&AI Lab's KULLM3, a Korean-English instruction-tuned model fine-tuned from SOLAR-10.7B-v1.0 on about 66,000 Korean and English instruction examples, requiring its fixed system prompt at inference. Provider: Korea University. Released 2024-04-02. Access: Open.
Aggregate Rating 1446 ± 18, rank #1022 of 1656 rated models, from 19 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-EQ-Bench | 43.54 | EQ-Bench score | 93 |
| Open Korean LLM Leaderboard | 41.85 | Average Score (%) | 83.7 |
| Open Ko-LLM Leaderboard | 41.85 | Average (%) | 83.6 |
| Open Korean LLM - KorNAT-CKA | 30.09 | Normalized accuracy (%) | 81.8 |
| Open Korean LLM - Ko-IFEval | 39.52 | Strict accuracy, prompt/instruction mean (%) | 81.1 |
| Open Korean LLM - Ko-Winogrande | 64.48 | Accuracy (%) | 79 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 35.41 | Exact match, flexible extract (%) | 68.7 |
| LogicKor - Writing | 8.28 | Score (0-10) | 60.7 |
| Open Korean LLM - KorNAT-Harmless | 63.31 | Normalized accuracy (%) | 60 |
| LogicKor - Single-Turn | 6.69 | Score (0-10) | 57.8 |
| LogicKor - Coding | 7.28 | Score (0-10) | 50.2 |
| LogicKor - Understanding | 6.85 | Score (0-10) | 38.9 |
Interactive version: theaggregate.ai/model?slug=kullm3 · How It Works · Data refreshed daily, snapshot 2026-10-20.