The query.dev for sparse dataset is incorrect. As none of the embedding has value at the same dimension as of train embedding. Hence, the dot product is always coming out as zero.
if the dot product is zero between train and test vector then ground truth is also incorrect.
I tested it with small train dataset.
The query.dev for sparse dataset is incorrect. As none of the embedding has value at the same dimension as of train embedding. Hence, the dot product is always coming out as zero.
if the dot product is zero between train and test vector then ground truth is also incorrect.
I tested it with small train dataset.