Multilingual-Transfer
Collection
Pretraining models to find what allows multilingual transfer • 26 items • Updated • 2
LLaMA-3 7B pretrained on English-Arabic bilingual data, ~42k steps.
Final checkpoint: step 41,416
| Validation set | Cross-entropy loss (nats) | Perplexity |
|---|---|---|
| English (en) | 2.5379 | 12.65 |
AR (ar) |
3.3761 | 29.26 |
Note: Validation losses measured at step 41,416 (last validate before step-42000 checkpoint).
EEE-format evaluation results are stored under eval_results/eee/ in this repository.
Tasks: Global MMLU (EN/AR/RU), PIQA, ECLeKTic, Fictive Entity (2-rate mix).
Part of the The-CoLab multilingual-transfer collection.