Datasets:

---

Francesco-A

/

github-issues_huggingface-datasets

like 0

[ { "color": "a2eeef", "default": true, "description": "New feature or request", "id": 1935892871, "name": "enhancement", "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement" }, { "color": "E5583E", "default": false, "description": "Related to the dataset viewer on huggingface.co", "id": 3470211881, "name": "dataset-viewer", "node_id": "LA_kwDODunzps7O1zsp", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset-viewer" } ]

[ "Yes, I agree this feature is much needed. We could do something similar to what TF does (https://www.tensorflow.org/api_docs/python/tf/data/Dataset#cache). \r\n\r\nIdeally, if the entire streamed dataset is consumed/cached, the generated cache should be reusable for the Arrow dataset." ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "<details>\n<summary>Show benchmarks</summary>\n\nPyArrow==6.0.0\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.014525 / 0.011353 (0.003172) | 0.006871 / 0.011008 (-0.004137) | 0.135577 / 0.038508 (0.097069) | 0.039620 / 0.023109 (0.016511) | 0.499829 / 0.275898 (0.223931) | 0.571000 / 0.323480 (0.247520) | 0.009726 / 0.007986 (0.001740) | 0.005654 / 0.004328 (0.001325) | 0.104732 / 0.004250 (0.100482) | 0.046849 / 0.037052 (0.009796) | 0.486667 / 0.258489 (0.228178) | 0.543611 / 0.293841 (0.249770) | 0.056414 / 0.128546 (-0.072133) | 0.019974 / 0.075646 (-0.055672) | 0.484878 / 0.419271 (0.065606) | 0.059244 / 0.043533 (0.015711) | 0.490046 / 0.255139 (0.234907) | 0.517427 / 0.283200 (0.234227) | 0.114692 / 0.141683 (-0.026991) | 1.935935 / 1.452155 (0.483780) | 1.990253 / 1.492716 (0.497537) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.271008 / 0.018006 (0.253002) | 0.610964 / 0.000490 (0.610474) | 0.013423 / 0.000200 (0.013223) | 0.000523 / 0.000054 (0.000468) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.031940 / 0.037411 (-0.005472) | 0.130755 / 0.014526 (0.116229) | 0.146616 / 0.176557 (-0.029941) | 0.239386 / 0.737135 (-0.497749) | 0.146612 / 0.296338 (-0.149726) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.675383 / 0.215209 (0.460174) | 6.656828 / 2.077655 (4.579174) | 2.741231 / 1.504120 (1.237111) | 2.232921 / 1.541195 (0.691726) | 2.172116 / 1.468490 (0.703626) | 1.221623 / 4.584777 (-3.363154) | 5.683653 / 3.745712 (1.937941) | 5.344137 / 5.269862 (0.074275) | 2.969670 / 4.565676 (-1.596006) | 0.142107 / 0.424275 (-0.282168) | 0.015808 / 0.007607 (0.008201) | 0.767366 / 0.226044 (0.541321) | 8.059605 / 2.268929 (5.790676) | 3.333535 / 55.444624 (-52.111089) | 2.669619 / 6.876477 (-4.206857) | 2.652989 / 2.142072 (0.510917) | 1.526397 / 4.805227 (-3.278830) | 0.265609 / 6.500664 (-6.235055) | 0.082759 / 0.075469 (0.007290) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.631086 / 1.841788 (-0.210701) | 18.701351 / 8.074308 (10.627043) | 22.843802 / 10.191392 (12.652410) | 0.240134 / 0.680424 (-0.440290) | 0.046683 / 0.534201 (-0.487518) | 0.576488 / 0.579283 (-0.002795) | 0.650123 / 0.434364 (0.215759) | 0.661190 / 0.540337 (0.120853) | 0.759563 / 1.386936 (-0.627373) |\n\n</details>\nPyArrow==latest\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.009883 / 0.011353 (-0.001470) | 0.006692 / 0.011008 (-0.004316) | 0.098550 / 0.038508 (0.060042) | 0.035188 / 0.023109 (0.012078) | 0.463535 / 0.275898 (0.187637) | 0.472762 / 0.323480 (0.149282) | 0.007199 / 0.007986 (-0.000787) | 0.007961 / 0.004328 (0.003632) | 0.093140 / 0.004250 (0.088890) | 0.051752 / 0.037052 (0.014700) | 0.453412 / 0.258489 (0.194922) | 0.502741 / 0.293841 (0.208900) | 0.056006 / 0.128546 (-0.072540) | 0.020164 / 0.075646 (-0.055482) | 0.116828 / 0.419271 (-0.302444) | 0.067205 / 0.043533 (0.023672) | 0.442715 / 0.255139 (0.187576) | 0.472525 / 0.283200 (0.189326) | 0.122767 / 0.141683 (-0.018915) | 1.881366 / 1.452155 (0.429212) | 1.978786 / 1.492716 (0.486069) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.284180 / 0.018006 (0.266174) | 0.601556 / 0.000490 (0.601067) | 0.008455 / 0.000200 (0.008255) | 0.000112 / 0.000054 (0.000057) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.033515 / 0.037411 (-0.003896) | 0.136407 / 0.014526 (0.121881) | 0.143341 / 0.176557 (-0.033215) | 0.225394 / 0.737135 (-0.511741) | 0.153343 / 0.296338 (-0.142995) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.688202 / 0.215209 (0.472993) | 6.576502 / 2.077655 (4.498847) | 2.839175 / 1.504120 (1.335055) | 2.481152 / 1.541195 (0.939957) | 2.617227 / 1.468490 (1.148736) | 1.314854 / 4.584777 (-3.269922) | 5.805950 / 3.745712 (2.060238) | 3.188930 / 5.269862 (-2.080932) | 2.141719 / 4.565676 (-2.423957) | 0.145069 / 0.424275 (-0.279206) | 0.014567 / 0.007607 (0.006960) | 0.780000 / 0.226044 (0.553955) | 7.898016 / 2.268929 (5.629088) | 3.549060 / 55.444624 (-51.895564) | 2.856569 / 6.876477 (-4.019907) | 3.117719 / 2.142072 (0.975647) | 1.512560 / 4.805227 (-3.292668) | 0.262689 / 6.500664 (-6.237975) | 0.085979 / 0.075469 (0.010509) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.623550 / 1.841788 (-0.218238) | 19.597063 / 8.074308 (11.522755) | 21.293369 / 10.191392 (11.101977) | 0.263780 / 0.680424 (-0.416643) | 0.027289 / 0.534201 (-0.506912) | 0.560361 / 0.579283 (-0.018922) | 0.646288 / 0.434364 (0.211924) | 0.712699 / 0.540337 (0.172361) | 0.818332 / 1.386936 (-0.568604) |\n\n</details>\n</details>\n\n![](https://cml.dev/watermark.png#b304de5dde30c945ec1397d3b4fe86f3b323ca8b \"CML watermark\")\n" ]

[]

[]

[]

[ "Thank you @lhoestq for your comments! 😄 I added your suggested changes, ran the tests and regenerated `dataset_infos.json` and `dummy_data`." ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "<details>\n<summary>Show benchmarks</summary>\n\nPyArrow==8.0.0\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.008634 / 0.011353 (-0.002719) | 0.005147 / 0.011008 (-0.005861) | 0.102865 / 0.038508 (0.064357) | 0.080245 / 0.023109 (0.057136) | 0.401288 / 0.275898 (0.125390) | 0.419708 / 0.323480 (0.096228) | 0.006342 / 0.007986 (-0.001644) | 0.003998 / 0.004328 (-0.000330) | 0.078880 / 0.004250 (0.074630) | 0.068199 / 0.037052 (0.031147) | 0.389573 / 0.258489 (0.131084) | 0.417292 / 0.293841 (0.123451) | 0.048856 / 0.128546 (-0.079691) | 0.014165 / 0.075646 (-0.061481) | 0.348063 / 0.419271 (-0.071209) | 0.067547 / 0.043533 (0.024014) | 0.402251 / 0.255139 (0.147112) | 0.419478 / 0.283200 (0.136278) | 0.034846 / 0.141683 (-0.106837) | 1.773493 / 1.452155 (0.321338) | 1.930546 / 1.492716 (0.437830) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.211835 / 0.018006 (0.193829) | 0.545311 / 0.000490 (0.544821) | 0.006766 / 0.000200 (0.006566) | 0.000104 / 0.000054 (0.000050) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.035406 / 0.037411 (-0.002006) | 0.100769 / 0.014526 (0.086243) | 0.108667 / 0.176557 (-0.067890) | 0.193099 / 0.737135 (-0.544036) | 0.113539 / 0.296338 (-0.182799) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.586935 / 0.215209 (0.371726) | 5.895245 / 2.077655 (3.817591) | 2.528375 / 1.504120 (1.024255) | 2.228617 / 1.541195 (0.687423) | 2.295799 / 1.468490 (0.827309) | 0.859272 / 4.584777 (-3.725505) | 5.033434 / 3.745712 (1.287722) | 7.546587 / 5.269862 (2.276726) | 4.457137 / 4.565676 (-0.108539) | 0.099626 / 0.424275 (-0.324649) | 0.009296 / 0.007607 (0.001689) | 0.713498 / 0.226044 (0.487454) | 7.409385 / 2.268929 (5.140456) | 3.361418 / 55.444624 (-52.083206) | 2.681111 / 6.876477 (-4.195366) | 2.849598 / 2.142072 (0.707526) | 1.114863 / 4.805227 (-3.690364) | 0.215494 / 6.500664 (-6.285170) | 0.075807 / 0.075469 (0.000338) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.606458 / 1.841788 (-0.235330) | 23.751096 / 8.074308 (15.676788) | 21.279110 / 10.191392 (11.087718) | 0.220785 / 0.680424 (-0.459639) | 0.032688 / 0.534201 (-0.501513) | 0.530948 / 0.579283 (-0.048335) | 0.630056 / 0.434364 (0.195693) | 0.572743 / 0.540337 (0.032405) | 0.771853 / 1.386936 (-0.615083) |\n\n</details>\nPyArrow==latest\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.008693 / 0.011353 (-0.002660) | 0.004750 / 0.011008 (-0.006259) | 0.079764 / 0.038508 (0.041256) | 0.082096 / 0.023109 (0.058987) | 0.467198 / 0.275898 (0.191300) | 0.532361 / 0.323480 (0.208881) | 0.005836 / 0.007986 (-0.002149) | 0.004333 / 0.004328 (0.000005) | 0.080444 / 0.004250 (0.076194) | 0.065883 / 0.037052 (0.028831) | 0.464871 / 0.258489 (0.206382) | 0.575026 / 0.293841 (0.281185) | 0.057807 / 0.128546 (-0.070739) | 0.017462 / 0.075646 (-0.058185) | 0.093667 / 0.419271 (-0.325605) | 0.071466 / 0.043533 (0.027933) | 0.495846 / 0.255139 (0.240707) | 0.526100 / 0.283200 (0.242900) | 0.034852 / 0.141683 (-0.106831) | 1.884152 / 1.452155 (0.431998) | 1.922681 / 1.492716 (0.429965) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.250969 / 0.018006 (0.232963) | 0.504979 / 0.000490 (0.504489) | 0.000466 / 0.000200 (0.000266) | 0.000083 / 0.000054 (0.000028) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.032411 / 0.037411 (-0.005000) | 0.093184 / 0.014526 (0.078658) | 0.110798 / 0.176557 (-0.065759) | 0.165741 / 0.737135 (-0.571394) | 0.111022 / 0.296338 (-0.185317) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.661284 / 0.215209 (0.446075) | 6.622388 / 2.077655 (4.544733) | 3.095705 / 1.504120 (1.591585) | 2.745698 / 1.541195 (1.204503) | 2.694103 / 1.468490 (1.225612) | 0.862154 / 4.584777 (-3.722623) | 5.109985 / 3.745712 (1.364273) | 5.040362 / 5.269862 (-0.229499) | 3.072837 / 4.565676 (-1.492840) | 0.110421 / 0.424275 (-0.313854) | 0.008476 / 0.007607 (0.000869) | 0.910020 / 0.226044 (0.683975) | 8.123626 / 2.268929 (5.854698) | 3.813811 / 55.444624 (-51.630813) | 3.017244 / 6.876477 (-3.859232) | 3.061222 / 2.142072 (0.919150) | 1.073548 / 4.805227 (-3.731680) | 0.216327 / 6.500664 (-6.284338) | 0.072977 / 0.075469 (-0.002492) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.722482 / 1.841788 (-0.119305) | 23.706716 / 8.074308 (15.632407) | 23.192134 / 10.191392 (13.000742) | 0.276733 / 0.680424 (-0.403691) | 0.033538 / 0.534201 (-0.500663) | 0.602083 / 0.579283 (0.022799) | 0.578718 / 0.434364 (0.144354) | 0.558311 / 0.540337 (0.017974) | 0.740341 / 1.386936 (-0.646595) |\n\n</details>\n</details>\n\n![](https://cml.dev/watermark.png#7ac575b8ed57dac60d7ba33a616894f38601f84a \"CML watermark\")\n", "<details>\n<summary>Show benchmarks</summary>\n\nPyArrow==8.0.0\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.006862 / 0.011353 (-0.004491) | 0.004223 / 0.011008 (-0.006786) | 0.085931 / 0.038508 (0.047423) | 0.081437 / 0.023109 (0.058328) | 0.349542 / 0.275898 (0.073644) | 0.379881 / 0.323480 (0.056401) | 0.005651 / 0.007986 (-0.002334) | 0.003662 / 0.004328 (-0.000666) | 0.065251 / 0.004250 (0.061001) | 0.061599 / 0.037052 (0.024547) | 0.359681 / 0.258489 (0.101192) | 0.392502 / 0.293841 (0.098661) | 0.031300 / 0.128546 (-0.097246) | 0.008591 / 0.075646 (-0.067055) | 0.288577 / 0.419271 (-0.130694) | 0.062920 / 0.043533 (0.019388) | 0.348989 / 0.255139 (0.093850) | 0.362769 / 0.283200 (0.079569) | 0.030087 / 0.141683 (-0.111596) | 1.480748 / 1.452155 (0.028594) | 1.580413 / 1.492716 (0.087697) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.205804 / 0.018006 (0.187798) | 0.455386 / 0.000490 (0.454897) | 0.003134 / 0.000200 (0.002934) | 0.000077 / 0.000054 (0.000023) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.030252 / 0.037411 (-0.007159) | 0.087566 / 0.014526 (0.073041) | 0.098209 / 0.176557 (-0.078347) | 0.155816 / 0.737135 (-0.581319) | 0.098938 / 0.296338 (-0.197401) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.386688 / 0.215209 (0.171479) | 3.852777 / 2.077655 (1.775123) | 1.938688 / 1.504120 (0.434568) | 1.779234 / 1.541195 (0.238039) | 1.864262 / 1.468490 (0.395772) | 0.482472 / 4.584777 (-4.102305) | 3.658060 / 3.745712 (-0.087652) | 5.206489 / 5.269862 (-0.063373) | 3.262498 / 4.565676 (-1.303179) | 0.057523 / 0.424275 (-0.366752) | 0.007365 / 0.007607 (-0.000242) | 0.466886 / 0.226044 (0.240841) | 4.671026 / 2.268929 (2.402097) | 2.380357 / 55.444624 (-53.064268) | 2.096590 / 6.876477 (-4.779887) | 2.274415 / 2.142072 (0.132342) | 0.579705 / 4.805227 (-4.225522) | 0.134522 / 6.500664 (-6.366142) | 0.062232 / 0.075469 (-0.013237) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.245965 / 1.841788 (-0.595823) | 20.115180 / 8.074308 (12.040872) | 14.602983 / 10.191392 (4.411591) | 0.146890 / 0.680424 (-0.533533) | 0.018424 / 0.534201 (-0.515777) | 0.393941 / 0.579283 (-0.185342) | 0.413785 / 0.434364 (-0.020579) | 0.453344 / 0.540337 (-0.086993) | 0.655446 / 1.386936 (-0.731490) |\n\n</details>\nPyArrow==latest\n\n<details>\n<summary>Show updated benchmarks!</summary>\n\n### Benchmark: benchmark_array_xd.json\n\n| metric | read_batch_formatted_as_numpy after write_array2d | read_batch_formatted_as_numpy after write_flattened_sequence | read_batch_formatted_as_numpy after write_nested_sequence | read_batch_unformated after write_array2d | read_batch_unformated after write_flattened_sequence | read_batch_unformated after write_nested_sequence | read_col_formatted_as_numpy after write_array2d | read_col_formatted_as_numpy after write_flattened_sequence | read_col_formatted_as_numpy after write_nested_sequence | read_col_unformated after write_array2d | read_col_unformated after write_flattened_sequence | read_col_unformated after write_nested_sequence | read_formatted_as_numpy after write_array2d | read_formatted_as_numpy after write_flattened_sequence | read_formatted_as_numpy after write_nested_sequence | read_unformated after write_array2d | read_unformated after write_flattened_sequence | read_unformated after write_nested_sequence | write_array2d | write_flattened_sequence | write_nested_sequence |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.006807 / 0.011353 (-0.004546) | 0.004083 / 0.011008 (-0.006925) | 0.065389 / 0.038508 (0.026881) | 0.081056 / 0.023109 (0.057947) | 0.362823 / 0.275898 (0.086925) | 0.401928 / 0.323480 (0.078448) | 0.005452 / 0.007986 (-0.002533) | 0.003413 / 0.004328 (-0.000915) | 0.065238 / 0.004250 (0.060987) | 0.057264 / 0.037052 (0.020211) | 0.375713 / 0.258489 (0.117224) | 0.407858 / 0.293841 (0.114017) | 0.031580 / 0.128546 (-0.096966) | 0.008643 / 0.075646 (-0.067003) | 0.071693 / 0.419271 (-0.347578) | 0.049392 / 0.043533 (0.005859) | 0.370194 / 0.255139 (0.115055) | 0.384647 / 0.283200 (0.101447) | 0.024805 / 0.141683 (-0.116877) | 1.509511 / 1.452155 (0.057356) | 1.560193 / 1.492716 (0.067477) |\n\n### Benchmark: benchmark_getitem\\_100B.json\n\n| metric | get_batch_of\\_1024\\_random_rows | get_batch_of\\_1024\\_rows | get_first_row | get_last_row |\n|--------|---|---|---|---|\n| new / old (diff) | 0.234442 / 0.018006 (0.216436) | 0.458818 / 0.000490 (0.458329) | 0.000407 / 0.000200 (0.000207) | 0.000060 / 0.000054 (0.000006) |\n\n### Benchmark: benchmark_indices_mapping.json\n\n| metric | select | shard | shuffle | sort | train_test_split |\n|--------|---|---|---|---|---|\n| new / old (diff) | 0.031661 / 0.037411 (-0.005750) | 0.093143 / 0.014526 (0.078618) | 0.102205 / 0.176557 (-0.074352) | 0.155850 / 0.737135 (-0.581286) | 0.104345 / 0.296338 (-0.191994) |\n\n### Benchmark: benchmark_iterating.json\n\n| metric | read 5000 | read 50000 | read_batch 50000 10 | read_batch 50000 100 | read_batch 50000 1000 | read_formatted numpy 5000 | read_formatted pandas 5000 | read_formatted tensorflow 5000 | read_formatted torch 5000 | read_formatted_batch numpy 5000 10 | read_formatted_batch numpy 5000 1000 | shuffled read 5000 | shuffled read 50000 | shuffled read_batch 50000 10 | shuffled read_batch 50000 100 | shuffled read_batch 50000 1000 | shuffled read_formatted numpy 5000 | shuffled read_formatted_batch numpy 5000 10 | shuffled read_formatted_batch numpy 5000 1000 |\n|--------|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 0.419641 / 0.215209 (0.204432) | 4.200808 / 2.077655 (2.123153) | 2.218227 / 1.504120 (0.714107) | 2.052604 / 1.541195 (0.511409) | 2.150611 / 1.468490 (0.682121) | 0.482665 / 4.584777 (-4.102112) | 3.606541 / 3.745712 (-0.139172) | 3.310637 / 5.269862 (-1.959224) | 2.070200 / 4.565676 (-2.495476) | 0.056586 / 0.424275 (-0.367689) | 0.007826 / 0.007607 (0.000218) | 0.491037 / 0.226044 (0.264992) | 4.901538 / 2.268929 (2.632610) | 2.676402 / 55.444624 (-52.768223) | 2.363935 / 6.876477 (-4.512542) | 2.587813 / 2.142072 (0.445741) | 0.579302 / 4.805227 (-4.225926) | 0.132792 / 6.500664 (-6.367873) | 0.061865 / 0.075469 (-0.013604) |\n\n### Benchmark: benchmark_map_filter.json\n\n| metric | filter | map fast-tokenizer batched | map identity | map identity batched | map no-op batched | map no-op batched numpy | map no-op batched pandas | map no-op batched pytorch | map no-op batched tensorflow |\n|--------|---|---|---|---|---|---|---|---|---|\n| new / old (diff) | 1.354315 / 1.841788 (-0.487473) | 20.874516 / 8.074308 (12.800208) | 14.863559 / 10.191392 (4.672167) | 0.183635 / 0.680424 (-0.496789) | 0.018636 / 0.534201 (-0.515565) | 0.395317 / 0.579283 (-0.183966) | 0.410598 / 0.434364 (-0.023766) | 0.476485 / 0.540337 (-0.063853) | 0.643246 / 1.386936 (-0.743690) |\n\n</details>\n</details>\n\n![](https://cml.dev/watermark.png#4472a8795c603a95eef7c8f15cb04f1290cc8d11 \"CML watermark\")\n" ]

[ { "color": "d73a4a", "default": true, "description": "Something isn't working", "id": 1935892857, "name": "bug", "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug" } ]

[ "Hi ! Certain users might need these data (for training or simply to explore/index the dataset).\r\n\r\nFeel free to implement a map function that gets rid of these paragraphs and process the wikipedia dataset with it before training" ]

[]

[ "The CI error is unrelated to this PR\r\n\r\n... merging !" ]

[]

[]

[]

[ "Thanks!" ]

[]

[ "The CI failure is unrelated to this PR and fixed on master, merging :)" ]

[]

[]

[]

[ "Hi ! We plan to add streaming features in the future.\r\n\r\nThis should allow to load a dataset instantaneously without generating the arrow table. The trade-off is that accessing examples from a streaming dataset must be done in an iterative way, and with an additional (but hopefully minor) overhead.\r\nWhat do you think about this ?\r\n\r\nIf you have ideas or suggestions of what you expect from such features as a user, feel free to share them, this is really valuable to us !", "People mainly want this feature either because it takes too much time too make arrow tables, or they occupy too much memory on the disk. I think both the problem can be solved if we provide arrow tables themselves on datasets hub. Can we do this currently @lhoestq ? \r\n", "@lhoestq I think the ```try_from_hf_gcs``` provide the same functionality. What all datasets are available on HF GCS? Are all the datasets on huggingFace datasets hub are made available on GCS, automatically?", "Only datasets like wikipedia, wiki40b, wiki_dpr and natural questions are available already processed on the HF google storage. This is used to download directly the arrow file instead of building it from the original data files.", "@lhoestq How can we make sure that the data we upload on HuggingFace hub is available in form of preprocessed arrow files ?", "We're still working on this :) This will be available soon\r\nUsers will be able to put their processed arrow files on the Hub", "Hi! You can now use `Dataset.push_to_hub` to store preprocessed files on the Hub.\r\n\r\nAnd to avoid downloading preprocessed files, you can use streaming by setting `streaming=True` in `load_dataset`." ]

[]

[]

[]

[]

[]

[ "It's definitely going in the right direction ! Thanks for giving it a try\r\n\r\nI really like the API.\r\nIMO it's fine right now if we don't have all the dataset transforms (map, filter, etc.) as it can be done before building the multitask dataset, but it will be important to have them in the end.\r\nAll the formatting methods could easily be added though.\r\n\r\nI think there are some parts that will require some work with apache arrow like slicing. I can find a way to do it using pyarrow tables concatenation (I did something similar when implementing `__getitem__` with an input that is a list of indices [here](https://github.com/huggingface/nlp/pull/322/files#diff-73270df8d7f08c62a27e40806e1a5fb0R463-R469)). It is very fast and it allows to have the same output format as a normal Dataset.\r\n\r\nAlso maybe we should check that not only the columns but also the schemas match ?\r\nAnd maybe add the `seed` of the shuffling step as an argument ?\r\n\r\n", "Maybe we should remove the methods that are not implemented for now, WDYT @thomwolf ?", "That's an interesting first draft, thanks a lot for that and the user facing API is really nice.\r\n\r\nI think we should dive more into this and the questions of #217 before merging the first version though.\r\n\r\nIn particular, the typical way to do multi-tasking is usually to sample a task and then sample a batch within the selected task. I think we should probably stay be closer to this traditional approach, or at least make it very easy to do, rather than go to close to the T5 approach which is very specific to this paper.\r\n\r\nIn this regard, it seems important to find some way to address the remarks of @zphang. I'm still wondering if we should not adopt more of a sampling approach rather than an iteration approach.", "@thomwolf Thanks! I mainly wanted to get something working quickly for my own MTL research. I agree with a lot of the points you made so I'll convert this pull request back to a draft.\r\n\r\nFor your specific point about 'batch-level' multitask mixing, it would be a pretty trivial change to add a `batch_size` parameter and ensure every `batch_size` examples are from the same task. This would certainly work, but would add a notion of 'batches' to a Dataset, which does feel like a 'Sampler-level' concept and not a Dataset one. There's also the possibility of wanting some specific task-level sampling functionality (e.g. applying `SortishSampler` to each task) which would only work with this kind of 2 step sampling approach. My first proposal in the transformers repo was actually a Sampler https://github.com/huggingface/transformers/issues/4340. I wonder whether functionality at the sampler-level has a place in the vision for the `nlp` repo?\r\n\r\nI imagine following a sampling approach you'd have to abandon maintaining the same user-facing API as a standard dataset (A shame because replacing a single dataset seamlessly with a multitask one is a really nice user-experience).\r\n\r\nRandom half-Idea: You could have a class which accepts a list of any iterables (either a Dataset or a DataLoader which already is doing the batching). Not sure what interface you'd present though. hmmm. \r\n\r\nThere's definitely more discussion to have. \r\n", "Are there any updates on making multi-task learning more officially supported in the datasets/transformers libraries? \r\nGiven that many papers use more than one task, it would be great to have multi-task learning more officially supported and easier to use. There are a few notebooks/blogs about using HF Transformers for this, but they all mention that it's more of a hack and not really officially supported (e.g. [this notebook](https://colab.research.google.com/github/zphang/zphang.github.io/blob/master/files/notebooks/Multi_task_Training_with_Transformers_NLP.ipynb#scrollTo=xW8bnTgCsx5c), or [this blog](https://medium.com/@shahrukhx01/multi-task-learning-with-transformers-part-1-multi-prediction-heads-b7001cf014bf)). \r\n\r\n[jiant](https://github.com/nyu-mll/jiant) was a framework built on transformers that made multi-task learning a first class feature of the library until recently, but they stopped maintaining their library a month ago ([see here](https://github.com/nyu-mll/jiant)). \r\nThis could be a good reason to increase support from the HF team? @lhoestq @thomwolf \r\n\r\nI'm not advanced enough to contribute on this, but an up-to-date notebook showing how to train a model e.g. on both MLM and next-sentence-prediction would already be very useful!", "> Are there any updates on making multi-task learning more officially supported in the datasets/transformers libraries? Given that many papers use more than one task, it would be great to have multi-task learning more officially supported and easier to use. There are a few notebooks/blogs about using HF Transformers for this, but they all mention that it's more of a hack and not officially supported (e.g. [this notebook](https://colab.research.google.com/github/zphang/zphang.github.io/blob/master/files/notebooks/Multi_task_Training_with_Transformers_NLP.ipynb#scrollTo=xW8bnTgCsx5c), or [this blog](https://medium.com/@shahrukhx01/multi-task-learning-with-transformers-part-1-multi-prediction-heads-b7001cf014bf)).\r\n> \r\n> [jiant](https://github.com/nyu-mll/jiant) was a framework built on transformers that made multi-task learning a first class feature of the library until recently, but they stopped maintaining their library a month ago. This could be a good reason to increase support from the HF team? @lhoestq\r\n> \r\n> I'm not advanced enough to contribute on this, but an up-to-date notebook showing how to train a model e.g. on both MLM and NSP would already be very useful!\r\n\r\nI kinda stopped working on this as I didn't really get any response on an actual workable solution.\r\n\r\nThe problem that I came up against after initially being redirected here after [proposing this in the transformers repo](https://github.com/huggingface/transformers/issues/4340) ([among](https://github.com/huggingface/transformers/issues/6872) [others](https://github.com/huggingface/transformers/issues/1856)) , was the request be able to do the multitask mixing at the batch level as well as at the level of individual examples. As this repo doesn't really have the concept of 'batches' it would need to be implemented in the transformers repo, rather than here. You could then pick which level to do your multitask learning on.\r\n\r\nWork on T5 and as of last week, on [exT5](https://arxiv.org/pdf/2111.10952.pdf), have shown that multitask mixing on the example level works incredibly well (with a big enough batch size), so if you're ok doing that, then this pull request works.\r\n\r\nI completely agree that multitask learning is a vital part of modern NLP, nearly every piece of research code I write has at least some aspect of multitask learning (currently using this patch). Many of the top GLUE and SuperGLUE submissions are using some aspect of mutlitask learning. We need to support it.", "Fully agree. Batching and data loading is one important thing. The part I'm struggling with right now is the classification head (which is more part of the Transformers repo, but also essential for multi-task learning). @ghomasHudson, how do you tune two classification heads simultaneously? Say, when I want to fine-tune an existing base-model on some classification task (like NLI, or next-sentence-prediction) and at the same time add some MLM for regularisation & domain adaptation. In this case I need two classification heads, but I don't know how to switch them between the batches. ", "> Fully agree. Batching and data loading is one important thing. The part I'm struggling with right now is the classification head (which is more part of the Transformers repo, but also essential for multi-task learning). @ghomasHudson, how do you tune two classification heads simultaneously? Say, when I want to fine-tune an existing base-model on some classification task (like NLI, or next-sentence-prediction) and at the same time add some MLM for regularisation & domain adaptation. In this case I need two classification heads, but I don't know how to switch them between the batches.\r\n\r\nThis pull request is mainly focused on getting the data in the right format, but you're right that there's no easy way to pick between the heads without something like jiant. You could of course replicate this functionality yourself - probably by making a class that implements the functionality of both `ModelNameForSequenceClassification` or `ModelNameForMaskedLM` picking between them depending on some task parameter you add to the forward pass. \r\n\r\njiant make this approach model agnostic by [ignoring the custom per-model head implementations of huggingface](https://github.com/nyu-mll/jiant/blob/386d4e726a27becda1b03c241f064eb13c54860f/jiant/proj/main/modeling/heads.py#L17-L18), instead making generic versions. Then the jiant code [passes a `task` parameter](https://github.com/nyu-mll/jiant/blob/386d4e726a27becda1b03c241f064eb13c54860f/jiant/proj/main/modeling/primary.py#L107-L109) into their [JiantModel](https://github.com/nyu-mll/jiant/blob/386d4e726a27becda1b03c241f064eb13c54860f/jiant/proj/main/modeling/primary.py#L36-L79) wrapper. To implement this in huggingface transformers would require quite a few modifications to the current approach (potentially interfering with some other project aims e.g. code readability), so you might find it tricky to get a change like that accepted. It would be super cool though.\r\n\r\nAnd there's of course the exT5 way of doing things too where you sidestep this issue entirely by treating both tasks as text-to-text problems so you can end up with 100% shared parameters, e.g.\r\nMLM: `Lorem <mask_0> amet, consectetur <mask_1> do eiusmod tempor incididunt ut labore <mask_2>`\r\nNLI: `Premise: The Old One always comforted Ca'daan, except today. hypothesis: Ca'daan knew the Old One very well.`\r\nThis also allows you to do mixed batches of both tasks.\r\n\r\nPersonally, my research mainly focuses on this last approach, using the structure of the data itself to indicate the task rather than swapping in and out different parts of the network.", "Hi! `jiant` maintainer here, don't have much to add to the conversation yet but I'm happy to share my experience/thoughts on working with Multitask models if people have questions.", "Hi ! I think it could be easier to simply share as examples in `transformers` some code that uses `jiant` and/or subclass/reimplement some part of `transformers` for multitask ?", "> Hi ! I think it could be easier to simply share as examples in `transformers` some code that uses `jiant` and/or subclass/reimplement some part of `transformers` for multitask ?\r\n\r\nWell since `jiant` requires new huggingface models to be explicitly added (as there are [\"subtle differences in the models that jiant must abstract\"](https://github.com/nyu-mll/jiant/blob/master/guides/models/adding_models.md)), and isn't being maintained anymore, then the first option might be out of date quickly.\r\n\r\nIf `transformers` could move towards making the task-specific heads more generic and as well as [creating a new base model in the `__init__` method](https://github.com/huggingface/transformers/blob/43f953cc2eec804eba04e2a9ae164d1a33fd97a8/src/transformers/models/bert/modeling_bert.py#L1502), allowing it to be passed as an argument (along with other little tweaks to standardize the approach), then this functionality could be moved into `transformers` itself.\r\n\r\nIt does seem a little redundant to have `jiant` as a library abstracting all the idiosyncrasies of each model type, where this could be done directly in the `transformers` repo in a single place alongside the model.\r\n\r\nIt's not an easy problem to solve though, especially balanced with the desire to expose models with minimal abstraction. @zphang probably knows more about this than me though.", "As mentioned, one of the main obstacles is that HF/T doesn't support generic heads. At first glance, this should be easy, since the interface is quite simple: models output both a token-wise and a sequence representation (e.g. `[CLS]`), and heads use either one and output the corresponding predictions/losses.\r\n\r\nHowever, there are a number of cases where this doesn't work. One of them is multiple-choice tasks like HellaSwag, which is a multiple choice task with 4 text options. The way this is normally formatted is that you encode `context + question + option_X` for X=1..4, and then score all four options based on a scoring head and pick the highest scoring option as the prediction. This requires you to run the encoder on 4 separate inputs, which breaks the above abstraction (the task-specific model might need to call the encoder multiple times).\r\n\r\nAnother thing is batching. You can imagine with the above that you might want a different batch size for multiple-choice tasks compared to simpler classification tasks. This means you need task-specific batching as well. In addition, [it's been shown](https://arxiv.org/abs/2101.11038) that you really want to mix tasks within a single batch. This also leads into issues like how you want to sample different task examples, early stopping on them, how to mix the validation scores, etc. (`jiant` addressed these, through probably more-complicated-than-necessary configurations.)\r\n\r\nNone of these are insurmountable problems, but it requires some tweaking of the current code layout to get it to work. I would guess that it wouldn't take much work to get a 90% implementation.", "> Another thing is batching. You can imagine with the above that you might want a different batch size for multiple-choice tasks compared to simpler classification tasks. This means you need task-specific batching as well. In addition, [it's been shown](https://arxiv.org/abs/2101.11038) that you really want to mix tasks within a single batch. This also leads into issues like how you want to sample different task examples, early stopping on them, how to mix the validation scores, etc. (`jiant` addressed these, through probably more-complicated-than-necessary configurations.)\r\n\r\nThat's reassuring. exT5 find the same thing - that mixing tasks together in a batch gives better performance (provided the batch size is big enough that each batch contains a mix of different tasks). Assuming this, we can ignore doing things at the batch-level and just do this at the individual example level - in which case this pull request already does the data mixing part of the problem! Balancing different tasks could easily be added here by implementing temperature-scaled mixing, custom weights, etc...\r\n\r\nTo make a generic implementation of this using different heads would be hard (impossible?) without doing the sub-batching that Muppet do - in which case we're back at dealing with the 'batch' (sub-batch) level which would need an implementation in `transformers` not here.\r\n\r\n", "Mixing at example should work fine. One issue though is that, as mentioned above, different tasks maybe actually require different amounts of memory, so downstream the user would have to find some way to handle that. But this might be one of those \"the last/edge-case 10% is the hardest\" to handle kind of deals.", "Very true - there's always going to be those cases. I also feel that the way things are going, if we just leave this for a few years no one will be wanting to use task-specific heads anymore - it'll all be task prompts included in the input a-la GPT, T5, etc... which will make this substantially simpler to implement.\r\n\r\nIt's quite tricky to make a suitably non-opinionated generic version of this at the moment.", "> Is there an advantage to varying the proportions of each task in each batch\r\n\r\nSome tasks have much less data than others. E.g. SNLI vs. CoLA is almost a 100x difference, so people often sample differently-sized tasks differently.", "As a short-term solution, I like @lhoestq's suggestion to create a notebook that shows how to implement multi-task learning by subclassing some transformer & dataset classes in a general way. I've been trying to get @zphang's [great but old notebook](https://colab.research.google.com/github/zphang/zphang.github.io/blob/master/files/notebooks/Multi_task_Training_with_Transformers_NLP.ipynb#scrollTo=CQ39AbTAPAUi) on multi-task learning running today and I didn't get it to work, probably because it was implemented a long time ago with `transformers==2.11`, `torch==1.2`~ etc and installing older versions still caused errors.\r\nThere is also this [interesting new repo](https://github.com/shahrukhx01/multitask-learning-transformers), which has a cool way of enabling you to save and load a model with two classification heads ([see model here](https://huggingface.co/shahrukhx01/bert-multitask-query-classifiers) and blog post [here](https://medium.com/@shahrukhx01/multi-task-learning-with-transformers-part-1-multi-prediction-heads-b7001cf014bf)). Haven't tried it yet, but it only uses `BertForSequenceClassification` instead of the more general AutoModelForXYZ\r\n\r\n@zphang, would you maybe be up for contributing an updated version of your older notebook with the latest version of `transformers` and `datasets` which runs in today's colabs? I feel like this would be very helpful for the community and if you keep the classes/functions somewhat general, people can easily adapt it to their use cases! 🙏 :) \r\nWould be a great addition to the [HF notebooks](https://huggingface.co/docs/transformers/notebooks).\r\n\r\nIn the medium-term, I agree that it would be great to have more native support for this via the HF libraries. I feels weird that you can neither train the old BERT (trained on two tasks) nor any of the newer models, without some hacks. ", "@zphang would love to see the newer notebook as suggested by @MoritzLaurer " ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[]

[ "When I change from python3.6 to python3.8, it works! ", "Does it work when you install `nlp` from source on python 3.6?", "No, still the pickling error.", "I wasn't able to reproduce on google colab (python 3.6.9 as well) with \r\n\r\npickle==4.0\r\ndill=0.3.2\r\ntransformers==3.1.0\r\ndatasets=1.0.1 (also tried nlp 0.4.0)\r\n\r\nIf I try\r\n\r\n```python\r\nfrom datasets import load_dataset # or from nlp\r\nfrom transformers import BertTokenizer\r\n\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\ndataset = load_dataset(\"text\", data_files=file_path, split=\"train\")\r\ndataset = dataset.map(lambda ex: tokenizer(ex[\"text\"], add_special_tokens=True,\r\n truncation=True, max_length=512), batched=True)\r\ndataset.set_format(type='torch', columns=['input_ids'])\r\n```\r\nIt runs without error", "Closing since it looks like it's working on >= 3.6.9\r\nFeel free to re-open if you have other questions :)" ]

[]

[]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[ "Hi @patrickvonplaten - could you please advise whether this would be a welcomed change, and if so, who I consult regarding the unit-tests?", "I'd be happy with adding this change. @anton-l @lhoestq - what do you think?", "Cool ! I just fixed the tests by adding a dummy `validated.tsv` file in the dummy data archive of common_voice\r\n\r\nI wonder if you should separate the train/valid/test configuration from the validated/invalidated configuration of the splits ? \r\nIn particular having `validated` along with the train/valid/test splits could be a bit weird since it comprises them. We can do that if you think it makes more sense. Otherwise it's also good as it is right now :)\r\n", "Thanks! I think that there are 2 cases for using the validated partition: 1) trainset = {validated - dev - test}, dev and test as they come; 2) train, dev, and test sampled from validated manually with the desired ratios.\r\nIn either case, I think that it's quite a big change on the HF interface part, so could as well be taken care of in the client code. Or is it not? (In which case, what's the most compact way to implement this?)", "What's important IMO is to let the users as much flexibility as they need - so we try to not do too much regarding splits to not constrain users. So I guess the way it is right now is ok. Can you confirm that it's ok @patrickvonplaten and that it won't break some speech training script out there ?", "@lhoestq all split names are explicit in our example scripts, so this shouldn't break anything, feel free to merge :)\r\nI'll go ahead and add this to the official `mozilla-foundation` datasets as well ", "Good for me! This has no real down-sides IMO and surely won't break any training scripts." ]

[]

[]

[ { "color": "d876e3", "default": true, "description": "Further information is requested", "id": 1935892912, "name": "question", "node_id": "MDU6TGFiZWwxOTM1ODkyOTEy", "url": "https://api.github.com/repos/huggingface/datasets/labels/question" } ]

[ "In a dataset, labels correspond to the `ClassLabel` feature that has the `names` property that returns string represenation of the integer classes (or `num_classes` to get the number of different classes).", "I think the `features` attribute of the dataset object is what you are looking for:\r\n```\r\n>>> dataset.features\r\n{'sentence1': Value(dtype='string', id=None),\r\n 'sentence2': Value(dtype='string', id=None),\r\n 'label': ClassLabel(num_classes=2, names=['not_equivalent', 'equivalent'], names_file=None, id=None),\r\n 'idx': Value(dtype='int32', id=None)\r\n}\r\n>>> dataset.features[\"label\"].names\r\n['not_equivalent', 'equivalent']\r\n```\r\n\r\nFor reference: https://huggingface.co/docs/datasets/exploring.html" ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._", "@mishig25 is investigating why deprecated/added do not affect the enclosed text format when used in args docstring: no special formatting appears: \r\n- https://moon-ci-docs.huggingface.co/docs/datasets/pr_4429/en/package_reference/builder_classes#datasets.DatasetBuilder", "@albertvillanova please check now 👍 \r\nhttps://moon-ci-docs.huggingface.co/docs/datasets/pr_4429/en/package_reference/builder_classes#datasets.DatasetBuilder\r\n\r\n<img width=\"500\" alt=\"Screenshot 2022-06-06 at 10 20 34\" src=\"https://user-images.githubusercontent.com/11827707/172123471-fab97138-c903-4a71-ab7f-c90e5e43c58f.png\">\r\n", "Thanks @mishig25.\r\n\r\nJust one question: is it expected to have the deprecated box right edge not filling all the page width (contrary to the added box)?", "> Just one question: is it expected to have the deprecated box right edge not filling all the page width (contrary to the added box)?\r\n\r\nYes, that is expected 😊 because the depreacted box is being bounded by its parent box (the box for `name` argument in the screenshot above)" ]

[ { "color": "0e8a16", "default": false, "description": "Contribution to a dataset script", "id": 4564477500, "name": "dataset contribution", "node_id": "LA_kwDODunzps8AAAABEBBmPA", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20contribution" } ]

[ "I reflect your comment! @lhoestq ", "Wait, the format of the data just changed, so I'll take it into consideration and commit it.", "I update the code according to the dataset structure change.", "Thanks ! I think the dummy data are not valid yet - the dummy train.csv file only contains a partial example (the quotes `\"` start but never end).", "> Thanks ! I think the dummy data are not valid yet - the dummy train.csv file only contains a partial example (the quotes `\"` start but never end).\r\n\r\nHi! @lhoestq There is a problem. \r\n<img src=\"https://user-images.githubusercontent.com/42150335/149804142-3800e635-f5a0-44d9-9694-0c2b0c05f16b.png\" width=500>\r\n \r\nAs shown in the picture above, the conversation is divided into \"\\n\" in the \"conversion\" column. \r\nThat's why there's an error in the file path that only saved only five lines like below. \r\n\r\n```\r\n'idx', 'class', 'conversation'\r\n'0', '협박 대화', '\"지금 너 스스로를 죽여달라고 애원하는 것인가?'\r\n아닙니다. 죄송합니다.'\r\n죽을 거면 혼자 죽지 우리까지 사건에 휘말리게 해? 진짜 죽여버리고 싶게.'\r\n정말 잘못했습니다.\r\n```\r\n \r\nIn fact, these five lines are all one line. \r\n \r\n\r\n", "Hi ! I see, in this case ca you make sure that the dummy data has a full sample ?\r\n\r\nFeel free to open the dummy train.csv in the dummy_data.zip file and add the missing lines", "Sorry, I'm late to check! I'll send it to you soon!", "Thanks for your contribution, @sooftware. Are you still interested in adding this dataset?\r\n\r\nWe are removing the dataset scripts from this GitHub repo and moving them to the Hugging Face Hub: https://huggingface.co/datasets\r\n\r\nWe would suggest you create this dataset there, under this organization namespace: https://huggingface.co/tunib\r\n\r\nPlease, feel free to tell us if you need some help.", "Close this PR. Thanks!" ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_3836). All of your documentation changes will be reflected on that endpoint.", "Weird, the logo doesn't seem to be floating on my side (using Chrome) at https://huggingface.co/docs/datasets/master/en/index", "https://huggingface.co/docs/datasets/index\r\n\r\nThe needed css change from moon-landing just got deployed" ]

[ { "color": "d73a4a", "default": true, "description": "Something isn't working", "id": 1935892857, "name": "bug", "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug" } ]

[]

[]

[]

[]

[]

[ { "color": "e99695", "default": false, "description": "Requesting to add a new dataset", "id": 2067376369, "name": "dataset request", "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request" }, { "color": "bfdadc", "default": false, "description": "Vision datasets", "id": 3608941089, "name": "vision", "node_id": "LA_kwDODunzps7XHBIh", "url": "https://api.github.com/repos/huggingface/datasets/labels/vision" } ]

[ "Motivation: this can be a useful dataset for researchers working on climate change adaptation, urban studies, geography, etc. I'll see if I can figure out how to add it!" ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[]

[]

[ { "color": "cfd3d7", "default": true, "description": "This issue or pull request already exists", "id": 1935892865, "name": "duplicate", "node_id": "MDU6TGFiZWwxOTM1ODkyODY1", "url": "https://api.github.com/repos/huggingface/datasets/labels/duplicate" } ]

[ "A duplicate of https://github.com/huggingface/datasets/issues/5929" ]

[]

[ "I just tried to do the same but changing the `.wav` files to `.mp3` files and that doesn't fix it.", "I don't know if anyone will ever read this but I've tried to upload the same dataset with google colab and the output seems more clarifying. I didn't specify the train/test split so the dataset wasn't fully uploaded (or that is what I understood, might be wrong!!).\r\n\r\nNow, including the `drop_metadata` flag I can load the dataset normally (at least with colab notebook):\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"audiofolder\", data_dir=\"../archive/Dataset\", , drop_metadata=True)\r\n```\r\n\r\nI'll close the issue.", "@uriii3 Hello, I understand correctly that you converted your wav files to mp3?", "Yes but it didn't matter. I don't remember which of them I ended up working with." ]

[]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[ "Hi! Can you share one of the problematic audio files with us?\r\n\r\nI tried to reproduce the error with the following code: \r\n```python\r\nimport soundfile as sf\r\nimport numpy as np\r\nfrom datasets import Audio\r\n\r\nsf.write(\"empty.wav\", np.array([]), 16000)\r\nAudio(sampling_rate=24000).decode_example({\"path\": \"empty.wav\", \"bytes\": None})\r\n```\r\nBut without success.\r\n\r\nAlso, what version of `librosa` is installed in your env? (You can get this info with `python -c \"import librosa; print(librosa.__version__)`)\r\n\r\n" ]

[]

[ "Also this could be extended to the rest of the GitHub Action `yml` files, so let me know whether you want me to have a look into it! 🤗", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_5903). All of your documentation changes will be reflected on that endpoint." ]

[]

[ "@lhoestq I still have problems with processing `.opus` files with `soundfile` so I actually cannot fully check that it works but it should... Maybe this should be investigated in case of someone else would also have problems with that.\r\n\r\nAlso, as the data is in a private repo on the hub (before we come to a decision about audio data privacy), the needed checks cannot be done right now.", "@lhoestq I check the data redownloading for configs sharing the same languages, you were right: the data is downloaded once for each language. But samples are generated from scratch each time. Is it a supposed behavior? ", "> But samples are generated from scratch each time. Is it a supposed behavior?\r\n\r\nYea that's the way it works right now, because we generate one arrow file per configuration. Since changing the languages creates a new configuration, then it generates a new arrow file." ]

[]

[ "Hi @omar-araboghli, thanks for your proposal.\r\n\r\nHave you tried to use \"csv\" loader instead of \"text\"? That already has a `sep` argument.", "Hi @albertvillanova, thanks for the quick response!\r\n\r\nIndeed, I have been trying to use `csv` instead of `text`. However I am still not able to define range of rows as one sequence, that is achievable with passing `sample_by='paragraph'` to the `TextConfig`\r\n\r\nFor instance, the below code\r\n\r\n```python\r\nimport datasets\r\n\r\ndataset = datasets.load_dataset(\r\n path='csv',\r\n data_files={'train': TRAINING_SET_PATH},\r\n sep='\\t',\r\n header=None,\r\n column_names=['tokens', 'pos_tags', 'chunk_tags', 'ner_tags']\r\n)\r\n```\r\n\r\nleads to \r\n\r\n```python\r\ndataset\r\n>>> DatasetDict({\r\n train: Dataset({\r\n features: ['tokens', 'pos_tags', 'chunk_tags', 'ner_tags'],\r\n num_rows: 62543\r\n })\r\n})\r\n\r\ndataset['train'][0]\r\n>>> {'tokens': 'Distribution',\r\n 'pos_tags': 'NN',\r\n 'chunk_tags': 'O',\r\n 'ner_tags': 'O'\r\n}\r\n```\r\nIs there a way to deal with multiple csv rows as one dataset instance, where each column is a sequence of those rows ?" ]

[]

[ "@lhoestq added. " ]

[ { "color": "e99695", "default": false, "description": "Requesting to add a new dataset", "id": 2067376369, "name": "dataset request", "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request" }, { "color": "bfdadc", "default": false, "description": "Vision datasets", "id": 3608941089, "name": "vision", "node_id": "LA_kwDODunzps7XHBIh", "url": "https://api.github.com/repos/huggingface/datasets/labels/vision" } ]

[]

[]

[ "Maybe you can use nltk's treebank detokenizer ?\r\n```python\r\nfrom nltk.tokenize.treebank import TreebankWordDetokenizer\r\n\r\nTreebankWordDetokenizer().detokenize(\"it 's a charming and often affecting journey . \".split())\r\n# \"it's a charming and often affecting journey.\"\r\n```", "I am looking for alternative file URL here instead of adding extra processing code: https://github.com/huggingface/datasets/blob/171f2bba9dd8b92006b13cf076a5bf31d67d3e69/datasets/glue/glue.py#L174", "I don't know if there exists a detokenized version somewhere. Even the version on kaggle is tokenized" ]

[]

[]

[ { "color": "e99695", "default": false, "description": "Requesting to add a new dataset", "id": 2067376369, "name": "dataset request", "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request" } ]

[ "Sounds good, @mariamabarham do you want to give a look?\r\nI think we should have two configurations so we can allow either version of the dataset to be loaded with the `1.0` version being the default maybe.\r\n\r\nCc some authors of the great cos-e: @nazneenrajani @bmccann", "cos_e v1.0 is related to CQA v1.0 but only CQA v1.11 dataset is available on their website. Indeed their is lots of ids in cos_e v1, which are not in CQA v1.11 or the other way around.\r\n@sarahwie, @thomwolf, @nazneenrajani, @bmccann do you know where I can find CQA v1.0\r\n", "@mariamabarham I'm also not sure where to find CQA 1.0. Perhaps it's not possible to include this version of the dataset. I'll close the issue if that's the case.", "I do have a copy of the dataset. I can upload it to our repo.", "Great @nazneenrajani. let me know once done.\r\nThanks", "@mariamabarham @sarahwie I added them to the cos-e repo https://github.com/salesforce/cos-e/tree/master/data/v1.0", "You can now do\r\n```python\r\nfrom nlp import load_dataset\r\ncos_e = load_dataset(\"cos_e\", \"v1.0\")\r\n```\r\nThanks @mariamabarham !", "Thanks!", "@mariamabarham Just wanted to note that default behavior `cos_e = load_dataset(\"cos_e\")` now loads `v1.0`. Not sure if this is intentional (but the flag specification does work as intended). ", "> @mariamabarham Just wanted to note that default behavior `cos_e = load_dataset(\"cos_e\")` now loads `v1.0`. Not sure if this is intentional (but the flag specification does work as intended).\r\n\r\nIn the new version of `nlp`, if you try `cos_e = load_dataset(\"cos_e\")` it throws this error:\r\n```\r\nValueError: Config name is missing.\r\nPlease pick one among the available configs: ['v1.0', 'v1.11']\r\nExample of usage:\r\n\t`load_dataset('cos_e', 'v1.0')`\r\n```\r\nFor datasets with at least two configurations, we now force the user to pick one (no default)" ]

[ { "color": "d73a4a", "default": true, "description": "Something isn't working", "id": 1935892857, "name": "bug", "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug" } ]

[ "@patrickvonplaten Hi! I saw that you prepared this? :)", "Another thing, but maybe this should be a separate issue: As I see from the code, it would try to use up to 16 simultaneous downloads? This is problematic for Librispeech or anything on OpenSLR. On [the homepage](https://www.openslr.org/), it says:\r\n\r\n> If you want to download things from this site, please download them one at a time, and please don't use any fancy software-- just download things from your browser or use 'wget'. We have a firewall rule to drop connections from hosts with more than 5 simultaneous connections, and certain types of download software may activate this rule.\r\n\r\nRelated: https://github.com/tensorflow/datasets/issues/3885", "Hey @albertz,\r\n\r\nNice to see you here! It's been a while ;-) ", "Sorry maybe the docs haven't been super clear here. By `split` we mean one of `train.500`, `train.360`, `train.100`, `validation`, `test`. For Librispeech, you'll have to specific a config (either `other` or `clean`) though:\r\n\r\n```py\r\ndatasets.load_dataset(\"librispeech_asr\", \"clean\")\r\n```\r\n\r\nshould work and give you all splits (being \"train\", \"test\", ...) for the clean config of the dataset.\r\n", "If you need both `\"clean\"` and `\"other\"` I think you'll have to do concatenate them as follows: \r\n\r\n```py\r\nfrom datasets import concatenate_datasets, load_dataset\r\n\r\nother = load_dataset(\"librispeech_asr\", \"other\")\r\nclean = load_dataset(\"librispeech_asr\", \"clean\")\r\n\r\nlibrispeech = concatenate_datasets([other, clean])\r\n```\r\n\r\nSee https://huggingface.co/docs/datasets/v2.1.0/en/process#concatenate", "Downloading one split would be:\r\n\r\n```py\r\nfrom datasets import load_dataset\r\n\r\nother = load_dataset(\"librispeech_asr\", \"other\", split=\"train.500\")\r\n```\r\n\r\n\r\n", "cc @lhoestq FYI maybe the docs can be improved here", "Ah thanks. But wouldn't it be easier/nicer (and more canonical) to just make it in a way that simply `load_dataset(\"librispeech_asr\")` works?", "Pinging @lhoestq here, think this could make sense! Not sure however how the dictionary would then look like", "Would it make sense to have `clean` as the default config ?\r\n\r\nAlso I think `load_dataset(\"librispeech_asr\")` should have raised you an error that says that you need to specify a config\r\n\r\nI also opened a PR to improve the doc: https://github.com/huggingface/datasets/pull/4183", "> Would it make sense to have `clean` as the default config ?\r\n\r\nI think a user would expect that the default would give you the full dataset.\r\n\r\n> Also I think `load_dataset(\"librispeech_asr\")` should have raised you an error that says that you need to specify a config\r\n\r\nIt does raise an error, but this error confused me because I did not understand why I needed a config, or why I could not simply download the whole dataset, which is what people usually do with Librispeech.\r\n", "+1 for @albertz. Also think lots of people download the whole dataset (`\"clean\"` + `\"other\"`) for Librispeech.\r\n\r\nThink there are also some people though who:\r\n- a) Don't have the memory to store the whole dataset\r\n- b) Just want to evaluate on one of the two configs", "Ok ! Adding the \"all\" configuration would do the job then, thanks ! In the \"all\" configuration we can merge all the train.xxx splits into one \"train\" split, or keep them separate depending on what's the most practical to use (probably put everything in \"train\" no ?)", "I'm not too familiar with how to work with HuggingFace datasets, but people often do some curriculum learning scheme, where they start with train.100, later go over to train.100 + train.360, and then later use the whole train (960h). It would be good if this is easily possible.\r\n", "Hey @albertz, \r\n\r\nopened a PR here. Think by adding the \"subdataset\" class to each split \"train\", \"dev\", \"other\" as shown here: https://github.com/huggingface/datasets/pull/4184/files#r853272727 it should be easily possible (e.g. with the filter function https://huggingface.co/docs/datasets/v2.1.0/en/package_reference/main_classes#datasets.Dataset.filter )", "But also since everything is cached one could also just do:\r\n\r\n```python\r\nload_dataset(\"librispeech\", \"clean\", \"train.100\")\r\nload_dataset(\"librispeech\", \"clean\", \"train.100+train.360\")\r\nload_dataset(\"librispeech\" \"all\", \"train\") \r\n```", "Hi @patrickvonplaten ,\r\n\r\nload_dataset(\"librispeech_asr\", \"clean\", \"train.100\") actually downloads the whole dataset and not the 100 hr split, is this a bug?", "Hmm, I don't really see how that's possible: https://github.com/huggingface/datasets/blob/d22e39a0693d4be7410cf9a5d41fd5aac22be3cc/datasets/librispeech_asr/librispeech_asr.py#L51\r\n\r\nNote that all datasets related to `\"clean\"` are downloaded, but only `\"train.100\"` should be used. \r\n\r\ncc @lhoestq @albertvillanova @mariosasko can we do anything against download dataset links that are not related to the \"split\" that one actually needs. E.g. why should the split `\"train.360\"` be downloaded if for the user executes the above command:\r\n\r\n```py\r\nload_dataset(\"librispeech_asr\", \"clean\", \"train.100\")\r\n```", "@patrickvonplaten This problem is a bit harder than it may seem, and it has to do with how our scripts are structured - `_split_generators` downloads data for a split before its definition. There was an attempt to fix this in https://github.com/huggingface/datasets/pull/2249, but it wasn't flexible enough. Luckily, I have a plan of attack, and this issue is on our short-term roadmap, so I'll work on it soon.\r\n\r\nIn the meantime, one can use streaming or manually download a dataset script, remove unwanted splits and load a dataset via `load_dataset`.", "> load_dataset(\"librispeech_asr\", \"clean\", \"train.100\") actually downloads the whole dataset and not the 100 hr split, is this a bug?\r\n\r\nSince this bug is still there and google led me here when I was searching for a solution, I am writing down how to quickly fix it (as suggested by @mariosasko) for whoever else is not familiar with how the HF Hub works.\r\n\r\nDownload the [librispeech_asr.py](https://huggingface.co/datasets/librispeech_asr/blob/main/librispeech_asr.py) script and remove the unwanted splits both from the [`_DL_URLS` dictionary](https://huggingface.co/datasets/librispeech_asr/blob/main/librispeech_asr.py#L47-L68) and from the [`_split_generators` function](https://huggingface.co/datasets/librispeech_asr/blob/main/librispeech_asr.py#L121-L241).\r\n[Here ](https://huggingface.co/datasets/andreagasparini/librispeech_test_only) I made an example with only the test sets.\r\n\r\nThen either save the script locally and load the dataset via \r\n```python\r\nload_dataset(\"${local_path}/librispeech_asr.py\")\r\n```\r\n\r\nor [create a new dataset repo on the hub](https://huggingface.co/new-dataset) named \"librispeech_asr\" and upload the script there, then you can just run\r\n```python\r\nload_dataset(\"${hugging_face_username}/librispeech_asr\")\r\n```", "Fixed by https://github.com/huggingface/datasets/pull/4184" ]

[ { "color": "2edb81", "default": false, "description": "A bug in a dataset script provided in the library", "id": 2067388877, "name": "dataset bug", "node_id": "MDU6TGFiZWwyMDY3Mzg4ODc3", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20bug" } ]

[ "I couldn't reproduce on my side.\r\nIt looks like you were not able to generate all the examples, and you have the problem for each split train-test-validation.\r\nCould you try to enable logging, try again and send the logs ?\r\n```python\r\nimport logging\r\nlogging.basicConfig(level=logging.INFO)\r\n```", "here's the log\r\n```\r\n>>> import nlp\r\nimport logging\r\nlogging.basicConfig(level=logging.INFO)\r\nnlp.load_dataset('cnn_dailymail', '3.0.0')\r\n>>> import logging\r\n>>> logging.basicConfig(level=logging.INFO)\r\n>>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\nINFO:nlp.load:Checking /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py for additional imports.\r\nINFO:filelock:Lock 140443095301136 acquired on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\nINFO:nlp.load:Found main folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail\r\nINFO:nlp.load:Found specific version folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\nINFO:nlp.load:Found script file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.py\r\nINFO:nlp.load:Updating dataset infos file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/dataset_infos.json to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/dataset_infos.json\r\nINFO:nlp.load:Found metadata file for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.json\r\nINFO:filelock:Lock 140443095301136 released on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\nINFO:nlp.info:Loading Dataset Infos from /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\nINFO:nlp.builder:Generating dataset cnn_dailymail (/u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0)\r\nINFO:nlp.builder:Dataset not on Hf google storage. Downloading and preparing it from source\r\nDownloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\nINFO:nlp.utils.info_utils:All the checksums matched successfully.\r\nINFO:nlp.builder:Generating split train\r\nINFO:nlp.arrow_writer:Done writing 285161 examples in 1240618482 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-train.arrow.\r\nINFO:nlp.builder:Generating split validation\r\nINFO:nlp.arrow_writer:Done writing 13255 examples in 56637485 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-validation.arrow.\r\nINFO:nlp.builder:Generating split test\r\nINFO:nlp.arrow_writer:Done writing 11379 examples in 48931393 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-test.arrow.\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/load.py\", line 520, in load_dataset\r\n builder_instance.download_and_prepare(\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 431, in download_and_prepare\r\n self._download_and_prepare(\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 488, in _download_and_prepare\r\n verify_splits(self.info.splits, split_dict)\r\n File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/utils/info_utils.py\", line 70, in verify_splits\r\n raise NonMatchingSplitsSizesError(str(bad_splits))\r\nnlp.utils.info_utils.NonMatchingSplitsSizesError: [{'expected': SplitInfo(name='test', num_bytes=49424491, num_examples=11490, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='test', num_bytes=48931393, num_examples=11379, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='train', num_bytes=1249178681, num_examples=287113, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='train', num_bytes=1240618482, num_examples=285161, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='validation', num_bytes=57149241, num_examples=13368, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='validation', num_bytes=56637485, num_examples=13255, dataset_name='cnn_dailymail')}]\r\n```", "> here's the log\r\n> \r\n> ```\r\n> >>> import nlp\r\n> import logging\r\n> logging.basicConfig(level=logging.INFO)\r\n> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n> >>> import logging\r\n> >>> logging.basicConfig(level=logging.INFO)\r\n> >>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n> INFO:nlp.load:Checking /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py for additional imports.\r\n> INFO:filelock:Lock 140443095301136 acquired on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\n> INFO:nlp.load:Found main folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail\r\n> INFO:nlp.load:Found specific version folder for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\n> INFO:nlp.load:Found script file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.py\r\n> INFO:nlp.load:Updating dataset infos file from https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/dataset_infos.json to /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/dataset_infos.json\r\n> INFO:nlp.load:Found metadata file for dataset https://s3.amazonaws.com/datasets.huggingface.co/nlp/datasets/cnn_dailymail/cnn_dailymail.py at /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad/cnn_dailymail.json\r\n> INFO:filelock:Lock 140443095301136 released on /u/jm8wx/.cache/huggingface/datasets/720d2e20d8dc6d98f21195a39cc934bb41dd0a40b57ea3d323661a7c5d70522c.d44c2417f4e0fe938ede0a684dcbb1fa9b4789de22e8a99c43103d4b4c374b3b.py.lock\r\n> INFO:nlp.info:Loading Dataset Infos from /p/qdata/jm8wx/datasets/nlp/src/nlp/datasets/cnn_dailymail/9645e0bc96f647decf46541f6f4bef6936ee82ace653ac362bab03309a46d4ad\r\n> INFO:nlp.builder:Generating dataset cnn_dailymail (/u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0)\r\n> INFO:nlp.builder:Dataset not on Hf google storage. Downloading and preparing it from source\r\n> Downloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\n> INFO:nlp.utils.info_utils:All the checksums matched successfully.\r\n> INFO:nlp.builder:Generating split train\r\n> INFO:nlp.arrow_writer:Done writing 285161 examples in 1240618482 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-train.arrow.\r\n> INFO:nlp.builder:Generating split validation\r\n> INFO:nlp.arrow_writer:Done writing 13255 examples in 56637485 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-validation.arrow.\r\n> INFO:nlp.builder:Generating split test\r\n> INFO:nlp.arrow_writer:Done writing 11379 examples in 48931393 bytes /u/jm8wx/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0.incomplete/cnn_dailymail-test.arrow.\r\n> Traceback (most recent call last):\r\n> File \"<stdin>\", line 1, in <module>\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/load.py\", line 520, in load_dataset\r\n> builder_instance.download_and_prepare(\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 431, in download_and_prepare\r\n> self._download_and_prepare(\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/builder.py\", line 488, in _download_and_prepare\r\n> verify_splits(self.info.splits, split_dict)\r\n> File \"/p/qdata/jm8wx/datasets/nlp/src/nlp/utils/info_utils.py\", line 70, in verify_splits\r\n> raise NonMatchingSplitsSizesError(str(bad_splits))\r\n> nlp.utils.info_utils.NonMatchingSplitsSizesError: [{'expected': SplitInfo(name='test', num_bytes=49424491, num_examples=11490, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='test', num_bytes=48931393, num_examples=11379, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='train', num_bytes=1249178681, num_examples=287113, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='train', num_bytes=1240618482, num_examples=285161, dataset_name='cnn_dailymail')}, {'expected': SplitInfo(name='validation', num_bytes=57149241, num_examples=13368, dataset_name='cnn_dailymail'), 'recorded': SplitInfo(name='validation', num_bytes=56637485, num_examples=13255, dataset_name='cnn_dailymail')}]\r\n> ```\r\n\r\nWith `nlp == 0.3.0` version, I'm not able to reproduce this error on my side.\r\nWhich version are you using for reproducing your bug?\r\n\r\n```\r\n>> nlp.load_dataset('cnn_dailymail', '3.0.0')\r\n\r\n8.90k/8.90k [00:18<00:00, 486B/s]\r\n\r\nDownloading: 100%\r\n9.37k/9.37k [00:00<00:00, 234kB/s]\r\n\r\nDownloading and preparing dataset cnn_dailymail/3.0.0 (download: 558.32 MiB, generated: 1.26 GiB, total: 1.81 GiB) to /root/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0...\r\nDownloading:\r\n159M/? [00:09<00:00, 16.7MB/s]\r\n\r\nDownloading:\r\n376M/? [00:06<00:00, 62.6MB/s]\r\n\r\nDownloading:\r\n2.11M/? [00:06<00:00, 333kB/s]\r\n\r\nDownloading:\r\n46.4M/? [00:02<00:00, 18.4MB/s]\r\n\r\nDownloading:\r\n2.43M/? [00:00<00:00, 2.62MB/s]\r\n\r\nDataset cnn_dailymail downloaded and prepared to /root/.cache/huggingface/datasets/cnn_dailymail/3.0.0/3.0.0. Subsequent calls will reuse this data.\r\n{'test': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 11490),\r\n 'train': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 287113),\r\n 'validation': Dataset(schema: {'article': 'string', 'highlights': 'string'}, num_rows: 13368)}\r\n\r\n>> ...\r\n\r\n```", "In general if some examples are missing after processing (hence causing the `NonMatchingSplitsSizesError `), it is often due to either\r\n1) corrupted cached files\r\n2) decoding errors\r\n\r\nI just checked the dataset script for code that could lead to decoding errors but I couldn't find any. Before we try to dive more into the processing of the dataset, could you try to clear your cache ? Just to make sure that it isn't 1)", "Yes thanks for the support! I cleared out my cache folder and everything works fine now" ]

[ { "color": "d73a4a", "default": true, "description": "Something isn't working", "id": 1935892857, "name": "bug", "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug" } ]

[ "the bug code locate in ：\r\n if data_args.task_name is not None:\r\n # Downloading and loading a dataset from the hub.\r\n datasets = load_dataset(\"glue\", data_args.task_name, cache_dir=model_args.cache_dir)", "Hi @jinec,\r\n\r\nFrom time to time we get this kind of `ConnectionError` coming from the github.com website: https://raw.githubusercontent.com\r\n\r\nNormally, it should work if you wait a little and then retry.\r\n\r\nCould you please confirm if the problem persists?", "cannot connect，even by Web browser，please check that there is some problems。", "I can access https://raw.githubusercontent.com/huggingface/datasets/1.7.0/datasets/glue/glue.py without problem...", "> I can access https://raw.githubusercontent.com/huggingface/datasets/1.7.0/datasets/glue/glue.py without problem...\r\n\r\nI can not access https://raw.githubusercontent.com/huggingface/datasets either, I am in China", "Finally i can access it, by the superfast software. Thanks", "> Finally i can access it, by the superfast software. Thanks\r\n\r\nExcuse me, I have the same problem as you, could you please tell me how to solve it？" ]

[]

[ "I would also like to add a test, but am not sure whether it should go into `test_builder` (more natural imo) or `test_load` (which already contains a lot of the things I have to import to run my current testing setup). For reference, what I run to test that it works looks like:\r\n\r\n```\r\nimport os\r\nfrom pathlib import Path\r\nimport shutil\r\n\r\nimport datasets\r\nfrom datasets.builder import DatasetBuilder\r\nfrom datasets.features import Features, Value\r\n\r\nDATASET_LOADING_SCRIPT_NAME = \"__dummy_dataset1__\"\r\n\r\nDATASET_LOADING_SCRIPT_CODE = \"\"\"\r\nimport os\r\n\r\nimport datasets\r\nfrom datasets import DatasetInfo, Features, Split, SplitGenerator, Value\r\n\r\n\r\nclass __DummyDataset1__(datasets.GeneratorBasedBuilder):\r\n\r\n def _info(self) -> DatasetInfo:\r\n return DatasetInfo(features=Features({\"text\": Value(\"string\")}))\r\n\r\n def _split_generators(self, dl_manager):\r\n return [\r\n SplitGenerator(Split.TRAIN, gen_kwargs={\"filepaths\": [os.path.join(dl_manager.manual_dir, \"train1.txt\"), os.path.join(dl_manager.manual_dir, \"train2.txt\")]}),\r\n SplitGenerator(Split.TEST, gen_kwargs={\"filepaths\": [os.path.join(dl_manager.manual_dir, \"test.txt\")]}),\r\n ]\r\n\r\n def _generate_examples(self, filepaths, **kwargs):\r\n idx = 0\r\n for filepath in filepaths:\r\n with open(filepath, \"r\", encoding=\"utf-8\") as f:\r\n for line in f:\r\n yield idx, {\"text\": line.strip()}\r\n idx += 1\r\n\"\"\"\r\n\r\n\r\ndef dataset_loading_script_dir(tmp_path):\r\n script_name = DATASET_LOADING_SCRIPT_NAME\r\n script_dir = tmp_path / script_name\r\n script_dir.mkdir()\r\n script_path = script_dir / f\"{script_name}.py\"\r\n with open(script_path, \"w\") as f:\r\n f.write(DATASET_LOADING_SCRIPT_CODE)\r\n return str(script_dir)\r\n\r\n\r\ndef data_dir(tmp_path):\r\n data_dir = tmp_path / \"data_dir\"\r\n data_dir.mkdir()\r\n with open(data_dir / \"train1.txt\", \"w\") as f:\r\n f.write(\"foo\\n\" * 10)\r\n with open(data_dir / \"train2.txt\", \"w\") as f:\r\n f.write(\"foo\\n\" * 10)\r\n with open(data_dir / \"test.txt\", \"w\") as f:\r\n f.write(\"bar\\n\" * 10)\r\n return str(data_dir)\r\n\r\n\r\ndef load_dataset_builder_multiprocessed(tmp_path):\r\n builder = datasets.load_dataset_builder(\r\n os.path.join(dataset_loading_script_dir(tmp_path), DATASET_LOADING_SCRIPT_NAME + \".py\"),\r\n data_dir=data_dir(tmp_path),\r\n )\r\n assert isinstance(builder, DatasetBuilder)\r\n assert builder.name == DATASET_LOADING_SCRIPT_NAME\r\n assert builder.info.features == Features({\"text\": Value(\"string\")})\r\n builder.download_and_prepare(tmp_path / \"prepare_target\", max_shard_size=500, num_proc=2)\r\n\r\nif __name__ == \"__main__\":\r\n tmp_path = \"tmp\"\r\n if os.path.exists(tmp_path):\r\n raise FileExistsError(f\"path {tmp_path} already exists\")\r\n os.makedirs(tmp_path)\r\n try:\r\n load_dataset_builder_multiprocessed(Path(tmp_path))\r\n finally:\r\n # pass\r\n shutil.rmtree(tmp_path)\r\n```", "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_5107). All of your documentation changes will be reflected on that endpoint.", "Nice ! I think the test can go in `test_builder.py` :)", "I've added sharded arrow dataset loading. Two WIP items in the PR:\r\n- ~~Order is not conserved (it seems like the sharded files are read in the wrong order)~~\r\n- the tqdm for preparing the splits is wrong (it compares against the size of the whole split rather than against the size of the multiprocessing shard, but I am not sure how to access the latter)\r\n\r\nAlso `naming.filenames_for_dataset_split` is not very elegant imo.\r\n\r\n@lvwerra if you don't care about order, as I do, it's functional for now but I'd still quite like to get to the bottom of this.", "Found the ordering bug ! (`glob.glob` returning stuff in arbitrary order)", "I fixed the tqdm to be less misleading, but it can't tell where to stop. I am a bit hesitant to add a top-level tqdm (on the shard iterator) since for most intents it will do 0 -> N shards straight, but I am not sure what is the best way to present that info here.", "I'm continuing the PR :)", "Did a few changes:\r\n- make shards naming consistent:\r\n - use `{builder_name}-{split_name}.{file_format}` when there's only 1 shard\r\n - otherwise use `{builder_name}-{split_name}-{shard_idx:05d}-of-{num_shards:05d}.{file_format}`\r\n- update the reader to support reading several shards\r\n - added a new `shard_lengths` field in `SplitInfo` (FYI it is saved in `dataset_info.json` next to the shards as usual)\r\n - it's None when there's only 1 shard\r\n - otherwise it's a list of integers that correspond to the number of rows per shard\r\n - implemented partial reading to only memory map the required shards\r\n - e.g. when someone asks for a partial split like `train[:10%]`\r\n- align the sharding for beam datasets\r\n - no more combining into 1 big arrow file\r\n- added a tqdm bar\r\n - only one single bar, handled by the main process\r\n - gathers progress updates from other processes using `iflatmap_unordered`\r\n - shows the number of examples (even for datasets prepared by generating arrow tables)\r\n- disabled multiprocessing by default - users must pass `num_proc` explicitly\r\n- tests\r\n- docs", "Alright this is ready for review - sorry it ended up so big ^^'\r\n\r\nIf I can do anything to make it easier for your to review this PR @mariosasko let me know", "Multiprocessing is disabled by default but we may show a warning to encourage users to pass `num_proc` if the dataset is split in many files. Let me know what you think", "Hey, is this error seems to you guys natural? \r\n\r\nThe package built from `0d4e3907` commit tag, and here is the version displayed from the import ... \r\n```bash\r\n>>> datasets.__version__\r\n'2.6.1.dev0'\r\n>>> \r\n```\r\n\r\n```bash\r\n>>> data = load_dataset('dataset_loaders/rfw2latentplay', num_proc=14)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/somewhere//mambaforge/envs/datasets/lib/python3.8/site-packages/datasets/load.py\", line 1719, in load_dataset\r\n builder_instance = load_dataset_builder(\r\n File \"/somewhere//mambaforge/envs/datasets/lib/python3.8/site-packages/datasets/load.py\", line 1523, in load_dataset_builder\r\n builder_instance: DatasetBuilder = builder_cls(\r\n File \"/somewhere//mambaforge/envs/datasets/lib/python3.8/site-packages/datasets/builder.py\", line 1292, in __init__\r\n super().__init__(*args, **kwargs)\r\n File \"/somewhere//mambaforge/envs/datasets/lib/python3.8/site-packages/datasets/builder.py\", line 303, in __init__\r\n self.config, self.config_id = self._create_builder_config(\r\n File \"/somewhere//mambaforge/envs/datasets/lib/python3.8/site-packages/datasets/builder.py\", line 456, in _create_builder_config\r\n builder_config = self.BUILDER_CONFIG_CLASS(**config_kwargs)\r\nTypeError: __init__() got an unexpected keyword argument 'num_proc'\r\n```\r\n\r\nLet me know if I can help fixing this ... \r\n", "> Do we have some benchmarks to see the speed-up?\r\n\r\nOn my machine running `load_dataset(\"oscar-corpus/OSCAR-2201\", \"br\")` (which is split in shards) I go from 2-3k examples per sec to 4-5k examples per sec with num_proc=2 😉", "> Hey, is this error seems to you guys natural?\r\n>\r\n> The package built from 0d4e3907 commit tag, and here is the version displayed from the import ...\r\n\r\nI don't know where you got the `0d4e3907` commit tag from, it doesn't seem to be in this PR. You should try installing from this PR, or wait for it to be merged on `main`", "## Splits vs Shards\r\n\r\nMaybe it's a good idea to add some documentation on the `sharding` that can be achieved by passing `list` based arguments to the `SplitGenerator`s `gen_kwargs` ... \r\n\r\nI had to read the whole dataset generation source code to find this out ... \r\n\r\n\r\n", "> Maybe it's a good idea to add some documentation on the sharding that can be achieved by passing list based arguments to the SplitGenerators gen_kwargs ...\r\n\r\nThis is part of this PR :) you can check the changes in docs/source/dataset_script.mdx", "I took your comments into account @mariosasko thanks !\r\nLet me know if it's good for you now ;)", "The doc CI should be fixed by now hopefully, merging !" ]

[]

[]

[ { "color": "2edb81", "default": false, "description": "A bug in a dataset script provided in the library", "id": 2067388877, "name": "dataset bug", "node_id": "MDU6TGFiZWwyMDY3Mzg4ODc3", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20bug" } ]

[ "This was also mentioned in https://github.com/huggingface/datasets/issues/488 \r\n\r\nThe bucket where is data was stored seems to be unavailable now. Maybe we can change the URL to the ones in https://conferences.unite.un.org/uncorpus/en/downloadoverview ?", "Closing since this has been fixed by #1912" ]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_5214). All of your documentation changes will be reflected on that endpoint." ]

[ { "color": "e99695", "default": false, "description": "Requesting to add a new dataset", "id": 2067376369, "name": "dataset request", "node_id": "MDU6TGFiZWwyMDY3Mzc2MzY5", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset%20request" } ]

[ "Whoever handles this just needs to: \r\n\r\n- [ ] fork the HuggingFace Datasets repo\r\n- [ ] update the [existing dataset script](https://github.com/huggingface/datasets/blob/master/datasets/openslr/openslr.py) to add SLR25. Lots of copypasting from other sections of the script should make that easy. \r\nAmharic URL: https://www.openslr.org/resources/25/data_readspeech_am.tar.bz2. \r\nSwahili URL: https://www.openslr.org/resources/25/data_broadcastnews_sw.tar.bz2, \r\nWolof URL: https://www.openslr.org/resources/25/data_readspeech_wo.tar.bz2\r\n- [ ] update the [data card](https://github.com/huggingface/datasets/blob/master/datasets/openslr/README.md) to include information about SLR25. There's lots of other examples to draw from. \r\n- [ ] add the appropriate language tags to the data card as well. https://www.w3.org/International/questions/qa-choosing-language-tags, or just use `sw`, `am`, and `wo` for consistency. \r\n- [ ] make a pull request to merge your changes back into HuggingFace's repo", "... also the example in \"use in datasets library\" should be updated. It currently says \r\n\r\nBut you actually have to specify a subset, e.g. \r\n```python\r\ndataset = load_dataset(\"openslr\", \"SLR32\")\r\n```", "\r\n\r\n" ]

[]

[]

[ { "color": "E5583E", "default": false, "description": "Related to the dataset viewer on huggingface.co", "id": 3470211881, "name": "dataset-viewer", "node_id": "LA_kwDODunzps7O1zsp", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset-viewer" }, { "color": "51F745", "default": false, "description": "", "id": 4030248571, "name": "dataset-viewer-gated", "node_id": "LA_kwDODunzps7wOLZ7", "url": "https://api.github.com/repos/huggingface/datasets/labels/dataset-viewer-gated" } ]

[ "related (same dataset): https://github.com/huggingface/datasets/issues/4149. But the issue is different. Looking at it", "I thought this issue was related to the error I was seeing, but upon consideration I'd think the dataset viewer would return a 500 (unable to create the split like me) or a 404 (unable to load split b/c it was never created) error if it was having the issue I was seeing in #4149. 401 message makes it look like dataset viewer isn't passing through the identity of the user who has signed the licensing agreement when making the request to GET [examples.jsonl](https://huggingface.co/datasets/facebook/winoground/resolve/a86a60456fbbd242e9a744199071a6bd3e7fd9de/examples.jsonl).", "Pinging @SBrandeis, as it seems related to gated datasets and access tokens.", "To replicate:\r\n\r\n```python\r\n>>> import datasets\r\n>>> dataset= datasets.load_dataset('facebook/winoground', name='facebook--winoground', split='train', use_auth_token=\"hf_app_...\", streaming=True)\r\n>>> next(iter(dataset))\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 497, in __iter__\r\n for key, example in self._iter():\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 494, in _iter\r\n yield from ex_iterable\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 87, in __iter__\r\n yield from self.generate_examples_fn(**self.kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py\", line 439, in wrapper\r\n for key, table in generate_tables_fn(**kwargs):\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py\", line 85, in _generate_tables\r\n for file_idx, file in enumerate(files):\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/utils/streaming_download_manager.py\", line 679, in __iter__\r\n yield from self.generator(*self.args, **self.kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/utils/streaming_download_manager.py\", line 731, in _iter_from_urlpaths\r\n for dirpath, _, filenames in xwalk(urlpath, use_auth_token=use_auth_token):\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/datasets/utils/streaming_download_manager.py\", line 623, in xwalk\r\n for dirpath, dirnames, filenames in fs.walk(main_hop):\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/spec.py\", line 372, in walk\r\n listing = self.ls(path, detail=True, **kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 85, in wrapper\r\n return sync(self.loop, func, *args, **kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 65, in sync\r\n raise return_result\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/asyn.py\", line 25, in _runner\r\n result[0] = await coro\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 196, in _ls\r\n out = await self._ls_real(url, detail=detail, **kwargs)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 150, in _ls_real\r\n self._raise_not_found_for_status(r, url)\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/fsspec/implementations/http.py\", line 208, in _raise_not_found_for_status\r\n response.raise_for_status()\r\n File \"/home/slesage/hf/datasets-preview-backend/.venv/lib/python3.9/site-packages/aiohttp/client_reqrep.py\", line 1004, in raise_for_status\r\n raise ClientResponseError(\r\naiohttp.client_exceptions.ClientResponseError: 401, message='Unauthorized', url=URL('https://huggingface.co/datasets/facebook/winoground/resolve/a86a60456fbbd242e9a744199071a6bd3e7fd9de/examples.jsonl')\r\n```\r\n\r\n*edited to fix `use_token` -> `use_auth_token`, thx @odellus*", "~~Using your command to replicate and changing `use_token` to `use_auth_token` fixes the problem I was seeing in #4149.~~\r\nNevermind it gave me an iterator to a method returning the same 401s. Changing `use_token` to `use_auth_token` does not fix the issue.", "After investigation with @severo , we found a potential culprit: https://github.com/huggingface/datasets/blob/3cd0a009a43f9f174056d70bfa2ca32216181926/src/datasets/utils/streaming_download_manager.py#L610-L624\r\n\r\nThe streaming manager does not seem to pass `use_auth_token` to `fsspec` when streaming and not iterating content of a zip archive\r\n\r\ncc @albertvillanova @lhoestq ", "I was able to reproduce it on a private dataset, let me work on a fix", "Hey @lhoestq, Thanks for working on a fix! Any plans to merge #4173 into master? ", "Thanks for the heads up, I still need to fix some tests that are failing in the CI before merging ;)", "The fix has been merged, we'll do a new release soon, and update the dataset viewer", "Fixed, thanks!\r\n<img width=\"1119\" alt=\"Capture d’écran 2022-06-21 à 18 41 09\" src=\"https://user-images.githubusercontent.com/1676121/174853571-afb0749c-4178-4c89-ab40-bb162a449788.png\">\r\n" ]

[]

[]

[]

[ "Hi @cccntu you're right, as when `batch_examples={}` the current if-statement won't be triggered as the condition won't be satisfied, I'll prepare a PR to address it as well as add the regression tests so that this issue is handled properly.", "Hi @alvarobartt ,\r\nThanks for answering. Do you know when and why an empty batch is passed to this function? This only happened to me when processing with multiple workers, while chunking examples, I think.", "> Hi @alvarobartt , Thanks for answering. Do you know when and why an empty batch is passed to this function? This only happened to me when processing with multiple workers, while chunking examples, I think.\r\n\r\nSo it depends on how you're actually chunking the data as if you're not handling empty chunks `batch_examples={}` or `batch_examples=None`, you may end up running into this issue. So you could check the chunks before you actually call `ArrowWriter.write_batch`, but anyway the fix you proposed I think improves the logic of `write_batch` to avoid running into these issues.", "Thanks, I added a if-print and I found it does return an empty examples in the chunking function that is passed to `.map()`.", "Hi ! We consider an empty batch to look like this:\r\n```python\r\nempty_batch = {\r\n \"column_1\": [],\r\n \"column_2\": [],\r\n ...\r\n}\r\n```\r\n\r\nWhile `{}` corresponds to a batch with no columns.\r\n\r\nTherefore calling this code should fail, because the two batches don't have the same columns:\r\n```python\r\nwriter.write_batch({\"a\": [1, 2, 3]})\r\nwriter.write_batch({})\r\n```\r\n\r\nIf you want to write an empty batch, you should do this instead:\r\n```python\r\nwriter.write_batch({\"a\": [1, 2, 3]})\r\nwriter.write_batch({\"a\": []})\r\n```", "Makes sense, then the if-statement should remain the same or is it better to handle both cases separately using `if not batch_examples or len(next(iter(batch_examples.values()))) == 0: ...`?\r\n\r\nUpdating the regressions tests with an empty batch formatted as `{\"col_1\": [], \"col_2\": []}` instead of `{}` works fine with the current if, and also with the one proposed by @cccntu.", "> Makes sense, then the if-statement should remain the same or is it better to handle both cases separately using if not batch_examples or len(next(iter(batch_examples.values()))) == 0: ...?\r\n\r\nThere's a check later in the code that makes sure that the columns are the right ones, so I don't think we need to check for `{}` here\r\n\r\nIn particular the check `if not batch_examples or len(next(iter(batch_examples.values()))) == 0:` doesn't raise an error while it should, that why the old `if` is fine IMO\r\n\r\n> Updating the regressions tests with an empty batch formatted as {\"col_1\": [], \"col_2\": []} instead of {} works fine with the current if, and also with the one proposed by @cccntu.\r\n\r\nCool ! If you want you can update your PR to add the regression tests, to make sure that `{\"col_1\": [], \"col_2\": []}` works but not `{}`", "Great thanks for the response! So I'll just add that regression test and remove the current if-statement.", "Hi @lhoestq ,\r\n\r\nThanks for your explanation. Now I get it that `{}` means the columns are different. But wouldn't it be nice if the code can ignore it, like it ignores `{\"a\": []}`?\r\n\r\n\r\n--- \r\nBTW, \r\n> There's a check later in the code that makes sure that the columns are the right ones, so I don't think we need to check for {} here\r\n\r\nI remember the error happens around here:\r\nhttps://github.com/huggingface/datasets/blob/88a902d6474fae8d793542d57a4f3b0d187f3c5b/src/datasets/arrow_writer.py#L506-L507\r\nThe error says something like `arrays` and `schema` doesn't have the same length. And it's not very clear I passed a `{}`.\r\n\r\nedit: actual error message\r\n```\r\nFile \"site-packages/datasets/arrow_writer.py\", line 595, in write_batch\r\n pa_table = pa.Table.from_arrays(arrays, schema=schema)\r\n File \"pyarrow/table.pxi\", line 3557, in pyarrow.lib.Table.from_arrays\r\n File \"pyarrow/table.pxi\", line 1401, in pyarrow.lib._sanitize_arrays\r\nValueError: Schema and number of arrays unequal\r\n```", "> But wouldn't it be nice if the code can ignore it, like it ignores {\"a\": []}?\r\n\r\nI think it would make things confusing because it doesn't follow our definition of a batch: \"the columns of a batch = the keys of the dict\". It would probably break certain behaviors as well. For example if you remove all the columns of a dataset (using `.remove_colums(...)` or `.map(..., remove_columns=...)`), the writer has to write 0 columns, and currently the only way to tell the writer to do so using `write_batch` is to pass `{}`.\r\n\r\n> The error says something like arrays and schema doesn't have the same length. And it's not very clear I passed a {}.\r\n\r\nYea the message can actually be improved indeed, it's definitely not clear. Maybe we can add a line right before the call `pa.Table.from_arrays` to make sure the keys of the batch match the field names of the schema" ]

[]

[]

[]

[ "Thanks for the suggestions! I've included them and made a few minor tweaks along the way", "Please merge master into this branch to fix the CI, I just fixed metadata validation tests." ]

[]

[ "merging since the Ci is fixed on master" ]

[ { "color": "0075ca", "default": true, "description": "Improvements or additions to documentation", "id": 1935892861, "name": "documentation", "node_id": "MDU6TGFiZWwxOTM1ODkyODYx", "url": "https://api.github.com/repos/huggingface/datasets/labels/documentation" } ]

[ "_The documentation is not available anymore as the PR was closed or merged._" ]

[]

[]

[ { "color": "d73a4a", "default": true, "description": "Something isn't working", "id": 1935892857, "name": "bug", "node_id": "MDU6TGFiZWwxOTM1ODkyODU3", "url": "https://api.github.com/repos/huggingface/datasets/labels/bug" } ]

[ "Hi @jacobbieker, thanks for reporting!\r\n\r\nI have transferred this issue to our Hub team and they are investigating it. I keep you informed. ", "Hi @jacobbieker, we are investigating this issue on our side and we'll see if we can fix it, but please note that your repo is considered problematic for git. Here are the results of running https://github.com/github/git-sizer on it:\r\n\r\n```\r\nProcessing blobs: 147448 \r\nProcessing trees: 27 \r\nProcessing commits: 4 \r\nMatching commits to trees: 4 \r\nProcessing annotated tags: 0 \r\nProcessing references: 3 \r\n| Name | Value | Level of concern |\r\n| ---------------------------- | --------- | ------------------------------ |\r\n| Biggest objects | | |\r\n| * Trees | | |\r\n| * Maximum entries [1] | 167 k | !!!!!!!!!!!!!!!!!!!!!!!!!!!!!! |\r\n| | | |\r\n| Biggest checkouts | | |\r\n| * Number of files [2] | 189 k | *** |\r\n\r\n[1] aa057d2667c34c70c6146efc631f5c9917ff326e (refs/heads/main:2016.zarr/unknown)\r\n[2] 6897b7bf6440fdd16b2c39d08085a669e7eaa59d (refs/heads/main^{tree})\r\n```\r\n\r\nYou can check https://github.com/github/git-sizer for more information on how to avoid such pathological structures.", "Hi, thanks for getting back to me so quick! And yeah, I figured that was probably the problem. I was going to try to delete the repo, but couldn't through the website, so if that's the easiest way to solve it, I can regenerate the dataset in a different format with less tiny files, and you guys can delete the repo as it is. Zarr just saves everything as lots of small files to make chunks easy to load, which is why I was preferring that format, but maybne that just doesn't work well for HF datasets.", "Hi @jacobbieker,\r\n\r\nFor future use cases, our Hub team is still pondering whether to limit the maximum number of files per repo to avoid technical issues...\r\n\r\nOn the meantime, they have made a fix and your dataset is working: https://huggingface.co/datasets/openclimatefix/mrms" ]

[]