Cross-lingual Reasoning and Factual Knowledge on Global MMLU (test)
23.46Accuracy (RUS)w_TREX
Evaluation results
| Method | Links | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| w_TREXModel Size=200M, Training Volume=50B tokens, Tokenizer=TREX2026.01 | 23.46 | 23.75 | 23.09 | 23.69 | 24.17 | 26.75 | 22.95 | 24.12 | 23.99 | 23.17 | 23.56 | 22.93 | 22.95 | 24.51 | 23.01 | 25.14 | 23.83 | 0.59 | |
| w_llamaModel Size=200M, Training Volume=50B tokens, Tokenizer=LLaMA2026.01 | 22.95 | 23.1 | 22.91 | 24.16 | 22.89 | 22.95 | 22.92 | 22.97 | 23.31 | 23.05 | 22.9 | 22.97 | 22.97 | 24.09 | 22.92 | 24.72 | 23.24 | — |