Four-Hop Training?

[okhat]

Hi! Thanks for the great paper and for the code release!

I'm curious if it's straightforward to train a three- or four-hop retriever using this code. I can create something like hotpot_train_with_neg_v0.json and, instead of one bridge, have a list of 2--3 bridges for each example.

I see that I need to modify only mhop_dataset.py (which is used from train_mhop.py) so it uses the modified bridges. Do you expect this to be a simple modification? It seems that some of the remaining code also assumes at most one bridge?

[xwhan]

Hi Omar, it should be easy to modify the code the support more hops. Besides the mhop_dataset.py, you will need to modify

• the model to encode the query/target in multiple hops
• loss function and validation metrics
• validation code snippets in train_mhop.py

[okhat]

Thank you Wenhan!

It seems a little more involved than I hoped, unfortunately. I will close this issue for now; will reopen if necessary.