Low-resource language evaluation on MiLiC-Eval
30.4BOD ScoreTRIMIX (PPL)
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| TRIMIX (PPL)Model Family=Llama2, Selection Strategy=Perplexity-Guided (PPL), Backbone=7B-cpt + 13B-ins2026.04 | 30.4 | 33.2 | 13.8 | 46.1 | 50.8 | 43.2 | 53.6 | 38.7 | |
| 7B-cptModel Family=Llama2, Size=7B, Type=Continued Pre-trained2026.04 | 27.6 | 26.8 | 12.6 | 41.1 | 44.5 | 36.3 | 46.4 | 33.6 | |
| Proxy TuningModel Family=Llama2, Backbone=7B-cpt + 13B-ins2026.04 | 25.7 | 26.8 | 13.3 | 33.2 | 39.6 | 33.3 | 47.3 | 31.3 | |
| 13B-insModel Family=Llama2, Size=13B, Type=Instruction-tuned2026.04 | 22.1 | 22.7 | 16.4 | 25.7 | 19 | 22.7 | 34 | 23.2 | |
| 7B-baseModel Family=Llama2, Size=7B2026.04 | 17 | 18.3 | 12.9 | 26.4 | 21.8 | 19.3 | 31.7 | 21.1 |