ST, MT and UniIR training partition in Table 6

[lcxrocks]

Thank you for your excellent work!
In Table 6, there are 3 different fine-tuning settings: single-task (ST), multi-task (MT) and UniIR. According to your explanation in section 4.1, ST is fine-tuned on each specific dataset only. At the same time, MT and UniIR is finetuned on all M-BEIR training data. In addition, Table 6 is the evaluation result on M-BEIR_local, which is a task-specific pool provided by each original dataset.

While the purpose of Table 6 is to highlight the advantages of UniIR fine-tuning, could the observed performance gap (e.g., UniIR vs. ST and MT vs. ST) be attributed to inconsistencies in the training partition? Specifically, could the performance improvement benefit from fine-tuning on more training instances? Or is this concern unnecessary or meaningless?