This is triplet preference data used by ALMA-R model.
The triplet preference data, supporting 10 translation directions, is built upon the FLORES-200 development and test data. For each direction, we provide a source sentence along with three translations: one from GPT-4, another from ALMA-13B-LoRA, and a reference translation. For instance, in the English-German pair, our data structure is as follows:
de: Original… See the full description on the dataset page:
https://huggingface.co/datasets/haoranxu/ALMA-R-Preference.
