Doesn't the generator need to call the retriever when training the model?

[liumc14]

I looked at the training data for training the generator and found that there were only instruction, output, input (empty), id, and dataset_name. But when training the generator, shouldn't it be retrieved when it is predicted that it needs to be retrieved? Where is the search process reflected?

self-rag/retrieval_lm/finetune.py

Line 511 in 1fcdc42

remove_columns=[name for name in raw_datasets["train"].column_names if name not in ["input_ids", "labels", "attention_mask"]],