Cross-lingual similarity search on FLORES X→Eng (devtest)
13.1xSIM++ LowSONAR-200
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SONAR-200Training languages=200, Number of parameters=1.3B, Zero-shot evaluation (New languages)=true2025.05 | 13.1 | 10.1 | 7.3 | 10.2 | 52.7 | |
| SONAR-75Training languages=63 known + 12 new, Number of parameters=1.3B2025.05 | 9 | 7.7 | 5.8 | 7.5 | 12.6 | |
| SONAR-63Training languages=63 known, Number of parameters=1.3B, Zero-shot evaluation (New languages)=true2025.05 | 8.8 | 7.7 | 5.8 | 7.5 | 42.9 | |
| charSONAR-75Training languages=63 known + 12 new, Number of parameters=1.3B2025.05 | 8.4 | 7.2 | 5.5 | 7 | 12.3 | |
| charSONAR-63Training languages=63 known, Number of parameters=1.3B, Zero-shot evaluation (New languages)=true2025.05 | 8.3 | 7.2 | 5.5 | 7 | 42.2 |