Low-resource language evaluation on MiLi-Eval
54.8BODTRIMIX
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TRIMIXSelection Strategy=PPL, Backbone=Gemma32026.04 | 54.8 | 55.9 | 51.8 | 29.4 | 48 | |
| 12B-insModel Size=12B, Stage=ins, Backbone=Gemma32026.04 | 49.7 | 57.6 | 50.8 | 24.1 | 45.6 | |
| Proxy TuningStrategy=Proxy Tuning, Backbone=Gemma32026.04 | 49.6 | 54.4 | 48.5 | 24.6 | 44.3 | |
| 4B-cptModel Size=4B, Stage=cpt, Backbone=Gemma32026.04 | 35.7 | 36 | 33.2 | 19.1 | 31 | |
| 4B-baseModel Size=4B, Stage=base, Backbone=Gemma32026.04 | 24.2 | 32 | 24.7 | 17.2 | 24.5 |