MADAR in training data
Thanks for open-sourcing this great model! I've been comparing the translation quality of multiple LLMs in English to Arabic dialects and Jais scores very well on MADAR dataset. Was this dataset used as a part of the training data? The BLEU scores would suggest it was, since Jais scores around 40-50 BLEU on some of the test sets, compared to ~15 BLEU for gpt-4.1. Do you have any suggestions for different test sets that could be used for a fair comparison?
Hi @cepin , Jais-2 was trained on the MADAR training and validation sets only. It was not trained on the MADAR test sets. We also ran our own evaluations on MADAR test data, and Jais-2 performs among the best models of its size on this benchmark. Please ensure that you are using the testing subset to evaluate Jais-2 performance.