This is the same dataset as https://huggingface.co/datasets/giuliadc/mlsum_de_train_10k, but the samples are ordered differently and have different ids:
samples with id newsroom-i with 0 < i < 9804 have sample[i]["text"].split(" ")) < 1200 (i.e. the text has less than 1500 words, if we consider a word everything that is separated from other words by a whitespace) samples with id newsroom-i with 9804 <= i < 10000 have sample[i]["text"].split(" ")) >= 1200