Machine Translation on German-Romansh translation varieties average (test)
58Win-rate vs Gemini 3 Pro (preview)Human reference
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Human referenceEvaluation protocol=Segment-level human evaluation, Evaluation level=Segment-level, Averaging method=Averaged across varieties2026.03 | 58 | — | 44 | 49 | |
| LR→HR augmentationEvaluation protocol=Segment-level human evaluation, Evaluation level=Segment-level, Averaging method=Averaged across varieties2026.03 | 53 | 21 | — | 20 | |
| LR→HR augmentation + dictionary promptingEvaluation protocol=Segment-level human evaluation, Evaluation level=Segment-level, Averaging method=Averaged across varieties2026.03 | 50 | 25 | 19 | — | |
| Gemini 3 Pro (preview)Evaluation protocol=Segment-level human evaluation, Evaluation level=Segment-level, Averaging method=Averaged across varieties2026.03 | — | 10 | 14 | 16 |