Support streaming mlsum dataset (#4574)

* Fix streaming for servers not supporting HTTP range requests

* Support streaming mlsum dataset

* Update metadata JSON

* Update dummy data

* Fix summarization task tag

* Unpin s3fs to allow fsspec>2021.08.01

* Revert workaround once fsspec>2021.08.01 is allowed

* Pin min fsspec version with fixed BlockSizeError

* Pin min versions for s3fs, aiobotocore, boto3, botocore

* Update compatible minimum requirements

Commit from https://github.com/huggingface/datasets/commit/612377be3fb306b1551dd5e0687f09ff2956d583

README.md

@@ -20,12 +20,13 @@ source_datasets:
 - extended|cnn_dailymail
 - original
 task_categories:
+- summarization
 - translation
 - text-classification
 task_ids:
+- news-articles-summarization
 - multi-class-classification
 - multi-label-classification
-- summarization
 - topic-classification
 paperswithcode_id: mlsum
 pretty_name: MLSUM

dataset_infos.json

@@ -1 +1 @@
-{"de": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. \nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish. \nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. \nWe report cross-lingual comparative analyses based on state-of-the-art systems. \nThese highlight existing biases which motivate the use of a multi-lingual dataset. \n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "supervised_keys": null, "builder_name": "mlsum", "config_name": "de", "version": {"version_str": "1.0.0", "description": null, "datasets_version_to_prepare": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 846960392, "num_examples": 220887, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 47119589, "num_examples": 11394, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 46847660, "num_examples": 10701, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_test.zip": {"num_bytes": 17741147, "checksum": "447e3b1839ab94d5700cc2aedc0b52521404865b2589656acc90a654ed0de4ff"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_train.zip": {"num_bytes": 311059697, "checksum": "88e788437bae48af6b3d18a554af4b2794cc6143a137df3f56daa91a37e3ea7e"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_val.zip": {"num_bytes": 17771216, "checksum": "732620c32e1d3f393ee3193f57f1217d8549499eb4906e144252aaab39aa910b"}}, "download_size": 346572060, "dataset_size": 940927641, "size_in_bytes": 1287499701}, "es": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. \nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish. \nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. \nWe report cross-lingual comparative analyses based on state-of-the-art systems. \nThese highlight existing biases which motivate the use of a multi-lingual dataset. \n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "supervised_keys": null, "builder_name": "mlsum", "config_name": "es", "version": {"version_str": "1.0.0", "description": null, "datasets_version_to_prepare": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 1214558950, "num_examples": 266367, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 50643448, "num_examples": 10358, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 71263713, "num_examples": 13920, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_test.zip": {"num_bytes": 27386169, "checksum": "177cfcf358bc4aa9bce2753b8e9de4f6eb41d2c30b1a99ef29d64e70537a1c0d"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_train.zip": {"num_bytes": 466443036, "checksum": "a01f4b4b873aa6cdeae15952a22ede2146734d0b60e7297470a35956507c863a"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_val.zip": {"num_bytes": 19483214, "checksum": "e38fce9950008ec4b48963692891c4c94d51a1e307286fb596e093aeb1230c92"}}, "download_size": 513312419, "dataset_size": 1336466111, "size_in_bytes": 1849778530}, "fr": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. \nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish. \nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. \nWe report cross-lingual comparative analyses based on state-of-the-art systems. \nThese highlight existing biases which motivate the use of a multi-lingual dataset. \n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "supervised_keys": null, "builder_name": "mlsum", "config_name": "fr", "version": {"version_str": "1.0.0", "description": null, "datasets_version_to_prepare": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 1471965974, "num_examples": 392902, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 70413260, "num_examples": 16059, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 69660336, "num_examples": 15828, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_test.zip": {"num_bytes": 26753725, "checksum": "7954f97de0f3839421e7c4aba38c72cc052771bc795cbca211ca2faea7d7d3b8"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_train.zip": {"num_bytes": 566354864, "checksum": "0ac483d3722219ca633c0f614622d64bc4c71e05f1bfb576a9e99b08de5801ba"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_val.zip": {"num_bytes": 26879418, "checksum": "755956e3ee4ae7c5da388286e7226a55bf5a0802482dee241899af319394300d"}}, "download_size": 619988007, "dataset_size": 1612039570, "size_in_bytes": 2232027577}, "ru": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. \nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish. \nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. \nWe report cross-lingual comparative analyses based on state-of-the-art systems. \nThese highlight existing biases which motivate the use of a multi-lingual dataset. \n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "supervised_keys": null, "builder_name": "mlsum", "config_name": "ru", "version": {"version_str": "1.0.0", "description": null, "datasets_version_to_prepare": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 257389569, "num_examples": 25556, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 9128521, "num_examples": 750, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 9656422, "num_examples": 757, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_test.zip": {"num_bytes": 3710826, "checksum": "769e009716f952cffb9cd6b722b8606cdd620e43a6559f1f05b1f0626d43b979"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_train.zip": {"num_bytes": 98998463, "checksum": "225cd04763d6d1d760d9819e9808b052cd6e45fa94ed5e1950c67fa7cf997a72"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_val.zip": {"num_bytes": 3506501, "checksum": "afd47b28ba023450669ed96029295027dc32dc8c71507e9c27748bb89a3924bb"}}, "download_size": 106215790, "dataset_size": 276174512, "size_in_bytes": 382390302}, "tu": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. \nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish. \nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. \nWe report cross-lingual comparative analyses based on state-of-the-art systems. \nThese highlight existing biases which motivate the use of a multi-lingual dataset. \n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "supervised_keys": null, "builder_name": "mlsum", "config_name": "tu", "version": {"version_str": "1.0.0", "description": null, "datasets_version_to_prepare": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 641623383, "num_examples": 249277, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 25530709, "num_examples": 11565, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 27830260, "num_examples": 12775, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_test.zip": {"num_bytes": 9525991, "checksum": "31ea869addf0dd483ef7446bdd1167b25a890d1ed1f9a48e74a67eebc731f463"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_train.zip": {"num_bytes": 229214838, "checksum": "cc8daf8a104fd362296276744d5a5b304c4e21aff7b4ffa761be2f3b7f531f8a"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_val.zip": {"num_bytes": 8751987, "checksum": "b39680a7dd648c970e03a7abf2c63b569c23f2b3f09dc0ae2fca8b7ad8f08b12"}}, "download_size": 247492816, "dataset_size": 694984352, "size_in_bytes": 942477168}}
+{"de": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset.\nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish.\nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.\nWe report cross-lingual comparative analyses based on state-of-the-art systems.\nThese highlight existing biases which motivate the use of a multi-lingual dataset.\n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "mlsum", "config_name": "de", "version": {"version_str": "1.0.0", "description": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 846959840, "num_examples": 220887, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 47119541, "num_examples": 11394, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 46847612, "num_examples": 10701, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_train.jsonl": {"num_bytes": 905470695, "checksum": "989ba44a56faa347d235a9e78877ed1376696012fedd880e1319d0dc484eab92"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_val.jsonl": {"num_bytes": 50325428, "checksum": "1adf4d0417e24ac06270bc433fae59b419c903cf8f261cf9df58ccac9cd64ab7"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/de_test.jsonl": {"num_bytes": 50018031, "checksum": "cc50a14c672d677a5571c2a3997dd25f8e389952fd8d4edeedb204578574f758"}}, "download_size": 1005814154, "post_processing_size": null, "dataset_size": 940926993, "size_in_bytes": 1946741147}, "es": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset.\nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish.\nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.\nWe report cross-lingual comparative analyses based on state-of-the-art systems.\nThese highlight existing biases which motivate the use of a multi-lingual dataset.\n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "mlsum", "config_name": "es", "version": {"version_str": "1.0.0", "description": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 1214558302, "num_examples": 266367, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 50643400, "num_examples": 10358, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 71263665, "num_examples": 13920, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_train.jsonl": {"num_bytes": 1323607428, "checksum": "416352aae3fb03e91ab2f0ffad78dc12008221344650780c45c9420cc965bc76"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_val.jsonl": {"num_bytes": 55106773, "checksum": "d869bd630943edaf17091ae042804066ae4f6acec2a91cfa57b3ef58c04d89fe"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/es_test.jsonl": {"num_bytes": 77496953, "checksum": "3d79b32b1abb84df7efb7a908ca072e6095b33a975204693812918a7b5440bcd"}}, "download_size": 1456211154, "post_processing_size": null, "dataset_size": 1336465367, "size_in_bytes": 2792676521}, "fr": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset.\nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish.\nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.\nWe report cross-lingual comparative analyses based on state-of-the-art systems.\nThese highlight existing biases which motivate the use of a multi-lingual dataset.\n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "mlsum", "config_name": "fr", "version": {"version_str": "1.0.0", "description": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 1471965014, "num_examples": 392902, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 70413212, "num_examples": 16059, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 69660288, "num_examples": 15828, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_train.jsonl": {"num_bytes": 1686967395, "checksum": "bd2df28b3df5e6152ac87772789e9e5c5d65c6a3bcd8f3c753d823e18f1ac10f"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_val.jsonl": {"num_bytes": 81755034, "checksum": "7b3f0f8666b318de4974b6ea62d2cb8b123e2ed927a6c7e54ae582313724d7a2"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/fr_test.jsonl": {"num_bytes": 80843135, "checksum": "07b9495357ce92d386cbff57aef6140befbedbc8137334c0a3b3d6d89d2ee3e2"}}, "download_size": 1849565564, "post_processing_size": null, "dataset_size": 1612038514, "size_in_bytes": 3461604078}, "ru": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset.\nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish.\nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.\nWe report cross-lingual comparative analyses based on state-of-the-art systems.\nThese highlight existing biases which motivate the use of a multi-lingual dataset.\n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "mlsum", "config_name": "ru", "version": {"version_str": "1.0.0", "description": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 257389497, "num_examples": 25556, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 9128497, "num_examples": 750, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 9656398, "num_examples": 757, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_train.jsonl": {"num_bytes": 714056813, "checksum": "f5073d94688405b5b0079e81d5e729e54184025e1cc8b9775258f27e62d92bcf"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_val.jsonl": {"num_bytes": 25345347, "checksum": "c19c30f75e9830f5026905d6e67b046e13b3bb7210dc15254c3fed5ef4c32424"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/ru_test.jsonl": {"num_bytes": 26823947, "checksum": "19e64d27a555605a75f774492f6525d65fa259db37af0fead8d4eb614d289e68"}}, "download_size": 766226107, "post_processing_size": null, "dataset_size": 276174392, "size_in_bytes": 1042400499}, "tu": {"description": "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset.\nObtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages -- namely, French, German, Spanish, Russian, Turkish.\nTogether with English newspapers from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community.\nWe report cross-lingual comparative analyses based on state-of-the-art systems.\nThese highlight existing biases which motivate the use of a multi-lingual dataset.\n", "citation": "@article{scialom2020mlsum,\n  title={MLSUM: The Multilingual Summarization Corpus},\n  author={Scialom, Thomas and Dray, Paul-Alexis and Lamprier, Sylvain and Piwowarski, Benjamin and Staiano, Jacopo},\n  journal={arXiv preprint arXiv:2004.14900},\n  year={2020}\n}\n", "homepage": "", "license": "", "features": {"text": {"dtype": "string", "id": null, "_type": "Value"}, "summary": {"dtype": "string", "id": null, "_type": "Value"}, "topic": {"dtype": "string", "id": null, "_type": "Value"}, "url": {"dtype": "string", "id": null, "_type": "Value"}, "title": {"dtype": "string", "id": null, "_type": "Value"}, "date": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "mlsum", "config_name": "tu", "version": {"version_str": "1.0.0", "description": null, "major": 1, "minor": 0, "patch": 0}, "splits": {"train": {"name": "train", "num_bytes": 641622783, "num_examples": 249277, "dataset_name": "mlsum"}, "validation": {"name": "validation", "num_bytes": 25530661, "num_examples": 11565, "dataset_name": "mlsum"}, "test": {"name": "test", "num_bytes": 27830212, "num_examples": 12775, "dataset_name": "mlsum"}}, "download_checksums": {"https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_train.jsonl": {"num_bytes": 870364512, "checksum": "51c4d951b147e62d9f725e59d2dbfefee33387e246cf8490ce827c98c6ca9489"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_val.jsonl": {"num_bytes": 34390982, "checksum": "d353e88d0875c6a8245f30c3920a9ce43d5e7b03cc2a77a45cf1b8f37a97161c"}, "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/tu_test.jsonl": {"num_bytes": 37553466, "checksum": "43eed825edc1799b482885f862162a4349dcc3739fe2b407a5259ba6dc6c2344"}}, "download_size": 942308960, "post_processing_size": null, "dataset_size": 694983656, "size_in_bytes": 1637292616}}

dummy/de/1.0.0/dummy_data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7cadc08524dacf9f056faba873e8fec41ea56917817a084f28abcbc0d30ffef4
-size 1624
+oid sha256:a23ab2c7ab6dcc7a95f2fedfba55b53e1bca8ff6d0bd931ad4e9770552756f57
+size 25711

dummy/es/1.0.0/dummy_data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:389de7907667221cda2ee9b7a8776202782a29f61bc967c4d492b3ad713311d3
-size 1624
+oid sha256:4709e09c79315dfbcafd55923b5e1942dfd332966f189613ff90ddba7a6726d0
+size 29055

dummy/fr/1.0.0/dummy_data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2320b620711a4e87c4199a2fd27946f0bfc030b74b9a1e7c50a617b65338cd83
-size 1624
+oid sha256:13c007fd024a235b5115592abf0ea4b11b6e74a7b8b33dc0d0c255ff344b49fe
+size 25457

dummy/ru/1.0.0/dummy_data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b108a64791dc81653da0b25d8cad78fc32ee9e592609de9cc2b8030af99a49a
-size 1624
+oid sha256:a2879090d9491dd44d0a7e06f378d8ab4dbde854eab8179e1bb46c45145d663b
+size 44217

dummy/tu/1.0.0/dummy_data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e340cb9e19b52325d0b206317d051ddd5b7eb9ae2edd6767667a636f78f3ab51
-size 1624
+oid sha256:a6592f282f2c665ffba9ab0c84de2ded84c02d8c55c230196a442e847d807e81
+size 10359

mlsum.py

@@ -1,5 +1,4 @@
 import json
-import os
 
 import datasets
 
@@ -20,7 +19,8 @@ Together with English newspapers from the popular CNN/Daily mail dataset, the co
 We report cross-lingual comparative analyses based on state-of-the-art systems.
 These highlight existing biases which motivate the use of a multi-lingual dataset.
 """
-_URL = "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM/"
+
+_URL = "https://gitlab.lip6.fr/scialom/mlsum_data/-/raw/master/MLSUM"
 _LANG = ["de", "es", "fr", "ru", "tu"]
 
 
@@ -65,49 +65,30 @@ class Mlsum(datasets.GeneratorBasedBuilder):
         # dl_manager is a datasets.download.DownloadManager that can be used to
         # download and extract URLs
 
-        lang = str(self.config.name)
+        lang = self.config.name
         urls_to_download = {
-            "test": _URL + lang + "_test.zip",
-            "train": _URL + lang + "_train.zip",
-            "validation": _URL + lang + "_val.zip",
+            "train": f"{_URL}/{lang}_train.jsonl",
+            "validation": f"{_URL}/{lang}_val.jsonl",
+            "test": f"{_URL}/{lang}_test.jsonl",
         }
-        downloaded_files = dl_manager.download_and_extract(urls_to_download)
+        downloaded_files = dl_manager.download(urls_to_download)
 
         return [
             datasets.SplitGenerator(
-                name=datasets.Split.TRAIN,
-                # These kwargs will be passed to _generate_examples
-                gen_kwargs={
-                    "filepath": os.path.join(downloaded_files["train"], lang + "_train.jsonl"),
-                    "lang": lang,
-                },
-            ),
-            datasets.SplitGenerator(
-                name=datasets.Split.VALIDATION,
-                # These kwargs will be passed to _generate_examples
+                name=split,
                 gen_kwargs={
-                    "filepath": os.path.join(downloaded_files["validation"], lang + "_val.jsonl"),
-                    "lang": lang,
+                    "filepath": downloaded_files[split],
                 },
-            ),
-            datasets.SplitGenerator(
-                name=datasets.Split.TEST,
-                # These kwargs will be passed to _generate_examples
-                gen_kwargs={
-                    "filepath": os.path.join(downloaded_files["test"], lang + "_test.jsonl"),
-                    "lang": lang,
-                },
-            ),
+            )
+            for split in [datasets.Split.TRAIN, datasets.Split.VALIDATION, datasets.Split.TEST]
         ]
 
-    def _generate_examples(self, filepath, lang):
+    def _generate_examples(self, filepath):
         """Yields examples."""
         with open(filepath, encoding="utf-8") as f:
-            i = 0
-            for line in f:
+            for id_, line in enumerate(f):
                 data = json.loads(line)
-                i += 1
-                yield i, {
+                yield id_, {
                     "text": data["text"],
                     "summary": data["summary"],
                     "topic": data["topic"],