Query structure for specific tasks

[Jasonqte]

Nice work! One question, I want to run this model on my own dataset, but the query logic of my downstream tasks mainly focus on is 1p, 2p, 3p, 4p, 5p. Can I only use the above five structures during training process? Does this have an impact on model performance?

[sshockwave]

Thank you! Yes, the model supports tuning toward specific query structures. This can be set in the configuration file. Here is an example, which is also how we boost the model performance.

[Jasonqte]

Thank you for your reply! Is that mean I should change the pretrain task from '1p 2p 3p 2i 3i' to '1p, 2p, 3p, 4p, 5p' or keep the original pretrain task and set a new finetune task like reasoning_5p which has reasoning_train_modes of '1p, 2p, 3p, 4p, 5p' and reasoning_test_modes of '5p'. Which one seems better?

[sshockwave]

Adding 5p sampling to the pre-train task might or might not improve the performance, which really depends on your dataset. Either way, I believe that a separate finetune task is best for datasets that are unseen during pre-training.