Response-level language alignment on OLA EN-Ko Complex
47.9EN Inst -> KO Content Pass RateOLMo 3.1 (SFT+DPO)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| OLMo 3.1 (SFT+DPO)CS-DPO=false2026.01 | 47.9 | 77.84 | |
| CS-DPOCS-DPO=true, Base Model=OLMo 3.1 (SFT+DPO)2026.01 | 43.77 | 83.96 |