Named Entity Recognition on MasakhaNER isiXhosa 2.0
89.9Macro F1AfroXLMR-large
Evaluation results
| Method | Links | |
|---|---|---|
| AfroXLMR-largeSize=0.550B, Type=Encoder-Only2026.03 | 89.9 | |
| AfroXLMR-baseSize=0.270B, Type=Encoder-Only2026.03 | 88.6 | |
| AfriBERTaSize=0.126B, Type=Encoder-Only2026.03 | 85 | |
| MzansiLMVariant=mono-masakhaner-ft, Size=0.125B, Evaluation Protocol=Monolingual task-specific finetuning2026.03 | 48.4 | |
| MzansiLMVariant=multi-masakhaner-ft, Size=0.125B, Evaluation Protocol=Multilingual task-specific finetuning2026.03 | 24.8 | |
| MzansiLMVariant=general-ft, Size=0.125B, Evaluation Protocol=Multi-task instruction finetuning2026.03 | 13.6 | |
| Llama-3.1-70B-InstructSize=70B, Type=Decoder-Only2026.03 | 8.2 | |
| AfroLlama-V1Size=8B, Type=Decoder-Only2026.03 | 5.6 | |
| InkubaLM-0.4BSize=0.4B, Type=Decoder-Only2026.03 | 0.1 | |
| MzansiLMVariant=Base 0-shot, Size=0.125B, Evaluation Protocol=no finetuning and no in-context examples2026.03 | 0 | |
| Aya-101Size=13B, Type=Encoder-Decoder2026.03 | 0 |