An idf is regular for every corpus, and accounts for your ratio of documents which include the word "this". With this case, Now we have a corpus of two documents and all of these contain the word "this".epoch. For this reason a Dataset.batch used right after Dataset.repeat will generate batches that