Dataset Information

#3
by mosama - opened

@MohamedRashad Can you also provide information about the dataset it was trained on and everything?

Navid AI org

The dataset itself is not important in my opinion, the method we used (style based GRPO reward functions) is far more important and it can work on any kind of dataset you make.

@MohamedRashad agreed that the style-based GRPO reward design is the part others can reuse. The dataset still matters for two practical reasons. Teams evaluating Yehia need to rule out overlap with Arabic benchmarks such as the OALL tasks. And enterprises in the region need to know the source and licence of the training data before they deploy.

Even a short data card (sources, size, language and dialect mix, licence) would make adoption much easier. @mosama

Sign up or log in to comment