Machine Translation on FLORES-200 pol_Latn (devtest)
23.9Tie Rate (%)GRPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GRPOModel Scale=600M, Experiment=A, Sample Size=50 source sentences, Evaluation Protocol=Human preference (one-sided binomial test)2026.05 | 23.9 | 0.41 | 34.8 | — | 0.75 | |
| Baseline NLLB-200Model Scale=600M, Decoding=beam search, Fine-tuning=none, Sample Size=50 source sentences, Evaluation Protocol=Human preference (one-sided binomial test)2026.05 | 23.9 | — | — | 41.3 | 0.75 |