The ablation results on xFlickr&CO

[Arthurizijar]

Regarding results provided in Table 3 for xFlickr&CO with the "-w/o parallel sentence pairs" setting, I find it incomprehensible.

As described in the paper, the "-w/o parallel sentence pairs" setting uses only CC3M with translated text during pre-training, which means that the model sees almost only 3 out of 7 languages on xFlickr&CO during pre-training, so how it achieves about 43% of the average Recall@1.

Could you please provide the specific performance of this setting on each language of xFlickr&CO ?
Or am I misunderstanding the settings? Thanks!