Span-level error detection on WMT25 Czech-German (official human annotations)
16.66PrecisionGemini-3-Flash
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Gemini-3-FlashError Filtering=Filtered2026.05 | 16.66 | 19.97 | 18.17 | |
| Gemini-3-FlashError Filtering=Unfiltered2026.05 | 16.53 | 20.23 | 18.19 | |
| Gemma-3-27b-itError Filtering=Filtered2026.05 | 11.24 | 9.62 | 10.37 | |
| Gemma-3-27b-itError Filtering=Unfiltered2026.05 | 11.12 | 9.86 | 10.45 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Filtered2026.05 | 6.87 | 18.41 | 10.01 | |
| EuroLLM-9B-InstructError Filtering=Filtered2026.05 | 6.59 | 8.49 | 7.42 | |
| EuroLLM-9B-InstructError Filtering=Unfiltered2026.05 | 6.21 | 9.03 | 7.36 | |
| Qwen3-VL-30B-A3B-InstructError Filtering=Unfiltered2026.05 | 4.55 | 20.39 | 7.43 |