Fix loading Hub datasets with CSV metadata file

[albertvillanova]

Currently, the reading of the metadata file infers the file extension (.jsonl or .csv) from the passed filename. However, downloaded files from the Hub don't have file extension. For example:

• the original file: hf://datasets/__DUMMY_TRANSFORMERS_USER__/test-dataset-5916a4-16977085077831/metadata.jsonl
• corresponds to the downloaded path: /tmp/pytest-of-username/pytest-46/cache/datasets/downloads/9f5374dbb470f711f6b89d66a5eec1f19cc96324b26bcbebe29138bda6cb20e6, which does not have extension

In the case where the metadata file does not have an extension, the reader assumes it is a JSONL file, thus the reported error when trying to read a CSV file as a JSONL one: ArrowInvalid: JSON parse error: Invalid value. in row 0

This behavior was introduced by:

• Add support for CSV metadata files to ImageFolder #4837

This PR extracts the metadata file extension from the original filename (instead of the downloaded one) and passes it as a parameter to the read_metadata function.

Fix #6315.

[github-actions]

Show benchmarks

PyArrow==8.0.0

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

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
new / old (diff)	0.008896 / 0.011353 (-0.002456)	0.005811 / 0.011008 (-0.005197)	0.108582 / 0.038508 (0.070074)	0.096509 / 0.023109 (0.073399)	0.481725 / 0.275898 (0.205827)	0.534743 / 0.323480 (0.211263)	0.005517 / 0.007986 (-0.002468)	0.006479 / 0.004328 (0.002151)	0.081313 / 0.004250 (0.077062)	0.063578 / 0.037052 (0.026525)	0.493977 / 0.258489 (0.235488)	0.551897 / 0.293841 (0.258056)	0.051835 / 0.128546 (-0.076711)	0.014105 / 0.075646 (-0.061541)	0.385866 / 0.419271 (-0.033405)	0.069131 / 0.043533 (0.025598)	0.484780 / 0.255139 (0.229641)	0.493221 / 0.283200 (0.210021)	0.039560 / 0.141683 (-0.102123)	1.782331 / 1.452155 (0.330176)	1.899193 / 1.492716 (0.406477)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.329978 / 0.018006 (0.311972)	0.600839 / 0.000490 (0.600349)	0.013187 / 0.000200 (0.012987)	0.000499 / 0.000054 (0.000444)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.031835 / 0.037411 (-0.005576)	0.103740 / 0.014526 (0.089214)	0.115875 / 0.176557 (-0.060681)	0.189880 / 0.737135 (-0.547255)	0.132614 / 0.296338 (-0.163725)

Benchmark: benchmark_iterating.json

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
new / old (diff)	0.596255 / 0.215209 (0.381046)	5.967993 / 2.077655 (3.890339)	2.612675 / 1.504120 (1.108555)	2.251461 / 1.541195 (0.710266)	2.308585 / 1.468490 (0.840095)	0.816516 / 4.584777 (-3.768261)	5.241791 / 3.745712 (1.496079)	4.680745 / 5.269862 (-0.589117)	2.997370 / 4.565676 (-1.568307)	0.098632 / 0.424275 (-0.325643)	0.010912 / 0.007607 (0.003305)	0.659092 / 0.226044 (0.433047)	6.825562 / 2.268929 (4.556634)	3.323844 / 55.444624 (-52.120780)	2.796203 / 6.876477 (-4.080274)	2.946994 / 2.142072 (0.804922)	1.002814 / 4.805227 (-3.802413)	0.202613 / 6.500664 (-6.298051)	0.072011 / 0.075469 (-0.003459)

Benchmark: benchmark_map_filter.json

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
new / old (diff)	1.613873 / 1.841788 (-0.227914)	24.500990 / 8.074308 (16.426682)	21.941599 / 10.191392 (11.750207)	0.214450 / 0.680424 (-0.465974)	0.031227 / 0.534201 (-0.502974)	0.498297 / 0.579283 (-0.080986)	0.597460 / 0.434364 (0.163096)	0.558152 / 0.540337 (0.017815)	0.789693 / 1.386936 (-0.597243)

PyArrow==latest

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

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
new / old (diff)	0.011299 / 0.011353 (-0.000053)	0.005103 / 0.011008 (-0.005905)	0.083161 / 0.038508 (0.044653)	0.094201 / 0.023109 (0.071092)	0.560457 / 0.275898 (0.284559)	0.590459 / 0.323480 (0.266980)	0.007059 / 0.007986 (-0.000926)	0.004418 / 0.004328 (0.000090)	0.081343 / 0.004250 (0.077093)	0.067069 / 0.037052 (0.030016)	0.538137 / 0.258489 (0.279648)	0.600416 / 0.293841 (0.306575)	0.049046 / 0.128546 (-0.079500)	0.014299 / 0.075646 (-0.061347)	0.093631 / 0.419271 (-0.325641)	0.062536 / 0.043533 (0.019003)	0.557238 / 0.255139 (0.302099)	0.571050 / 0.283200 (0.287850)	0.035881 / 0.141683 (-0.105802)	1.918487 / 1.452155 (0.466332)	2.013979 / 1.492716 (0.521263)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.400995 / 0.018006 (0.382989)	0.634898 / 0.000490 (0.634408)	0.041809 / 0.000200 (0.041609)	0.000279 / 0.000054 (0.000224)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.034160 / 0.037411 (-0.003251)	0.109996 / 0.014526 (0.095470)	0.124335 / 0.176557 (-0.052222)	0.188100 / 0.737135 (-0.549035)	0.135897 / 0.296338 (-0.160442)

Benchmark: benchmark_iterating.json

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
new / old (diff)	0.639751 / 0.215209 (0.424542)	6.403312 / 2.077655 (4.325657)	3.146453 / 1.504120 (1.642333)	2.840358 / 1.541195 (1.299164)	2.908667 / 1.468490 (1.440177)	0.818767 / 4.584777 (-3.766010)	5.416939 / 3.745712 (1.671227)	4.853498 / 5.269862 (-0.416364)	3.023526 / 4.565676 (-1.542150)	0.110850 / 0.424275 (-0.313425)	0.013103 / 0.007607 (0.005496)	0.799720 / 0.226044 (0.573676)	7.837704 / 2.268929 (5.568775)	4.016526 / 55.444624 (-51.428099)	3.338965 / 6.876477 (-3.537512)	3.715721 / 2.142072 (1.573648)	1.088340 / 4.805227 (-3.716887)	0.213610 / 6.500664 (-6.287054)	0.079244 / 0.075469 (0.003775)

Benchmark: benchmark_map_filter.json

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
new / old (diff)	1.833175 / 1.841788 (-0.008612)	25.307218 / 8.074308 (17.232910)	23.716075 / 10.191392 (13.524683)	0.259114 / 0.680424 (-0.421310)	0.035171 / 0.534201 (-0.499029)	0.530128 / 0.579283 (-0.049155)	0.651484 / 0.434364 (0.217120)	0.589414 / 0.540337 (0.049077)	0.862691 / 1.386936 (-0.524245)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[mariosasko]

Thanks!

Surprised this bug was not reported earlier.

[albertvillanova]

Me too, I thought the same... quite surprised... 😮

[github-actions]

Show benchmarks

PyArrow==8.0.0

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

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
new / old (diff)	0.006929 / 0.011353 (-0.004423)	0.004345 / 0.011008 (-0.006663)	0.085522 / 0.038508 (0.047014)	0.083380 / 0.023109 (0.060271)	0.310332 / 0.275898 (0.034434)	0.350525 / 0.323480 (0.027045)	0.004367 / 0.007986 (-0.003618)	0.005503 / 0.004328 (0.001175)	0.066311 / 0.004250 (0.062061)	0.059545 / 0.037052 (0.022492)	0.314090 / 0.258489 (0.055601)	0.366661 / 0.293841 (0.072821)	0.031581 / 0.128546 (-0.096965)	0.008852 / 0.075646 (-0.066794)	0.289312 / 0.419271 (-0.129960)	0.052960 / 0.043533 (0.009427)	0.308134 / 0.255139 (0.052995)	0.330342 / 0.283200 (0.047142)	0.026157 / 0.141683 (-0.115526)	1.488463 / 1.452155 (0.036308)	1.561441 / 1.492716 (0.068725)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.327735 / 0.018006 (0.309729)	0.568162 / 0.000490 (0.567672)	0.012097 / 0.000200 (0.011897)	0.000438 / 0.000054 (0.000383)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.029503 / 0.037411 (-0.007909)	0.084327 / 0.014526 (0.069801)	0.102065 / 0.176557 (-0.074492)	0.157392 / 0.737135 (-0.579744)	0.101428 / 0.296338 (-0.194910)

Benchmark: benchmark_iterating.json

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
new / old (diff)	0.386767 / 0.215209 (0.171558)	3.870757 / 2.077655 (1.793102)	1.870048 / 1.504120 (0.365928)	1.678221 / 1.541195 (0.137026)	1.799423 / 1.468490 (0.330933)	0.477718 / 4.584777 (-4.107059)	3.618351 / 3.745712 (-0.127361)	3.577921 / 5.269862 (-1.691941)	2.146217 / 4.565676 (-2.419459)	0.056290 / 0.424275 (-0.367985)	0.007378 / 0.007607 (-0.000229)	0.460678 / 0.226044 (0.234633)	4.606243 / 2.268929 (2.337314)	2.303460 / 55.444624 (-53.141164)	1.982662 / 6.876477 (-4.893814)	2.103891 / 2.142072 (-0.038182)	0.570700 / 4.805227 (-4.234527)	0.131747 / 6.500664 (-6.368918)	0.060915 / 0.075469 (-0.014554)

Benchmark: benchmark_map_filter.json

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
new / old (diff)	1.286364 / 1.841788 (-0.555424)	20.106330 / 8.074308 (12.032022)	14.780833 / 10.191392 (4.589441)	0.164301 / 0.680424 (-0.516123)	0.018730 / 0.534201 (-0.515471)	0.398530 / 0.579283 (-0.180754)	0.418084 / 0.434364 (-0.016280)	0.468735 / 0.540337 (-0.071602)	0.690122 / 1.386936 (-0.696814)

PyArrow==latest

Show updated benchmarks!

Benchmark: benchmark_array_xd.json

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
new / old (diff)	0.007262 / 0.011353 (-0.004091)	0.004228 / 0.011008 (-0.006780)	0.065866 / 0.038508 (0.027358)	0.096151 / 0.023109 (0.073042)	0.409352 / 0.275898 (0.133454)	0.441234 / 0.323480 (0.117754)	0.005946 / 0.007986 (-0.002039)	0.003630 / 0.004328 (-0.000698)	0.066271 / 0.004250 (0.062020)	0.061567 / 0.037052 (0.024515)	0.409097 / 0.258489 (0.150608)	0.447675 / 0.293841 (0.153834)	0.032804 / 0.128546 (-0.095743)	0.008793 / 0.075646 (-0.066853)	0.070790 / 0.419271 (-0.348482)	0.048650 / 0.043533 (0.005117)	0.411021 / 0.255139 (0.155882)	0.421398 / 0.283200 (0.138198)	0.025305 / 0.141683 (-0.116378)	1.494826 / 1.452155 (0.042671)	1.580441 / 1.492716 (0.087724)

Benchmark: benchmark_getitem_100B.json

metric	get_batch_of_1024_random_rows	get_batch_of_1024_rows	get_first_row	get_last_row
new / old (diff)	0.321871 / 0.018006 (0.303865)	0.526471 / 0.000490 (0.525982)	0.006913 / 0.000200 (0.006713)	0.000108 / 0.000054 (0.000054)

Benchmark: benchmark_indices_mapping.json

metric	select	shard	shuffle	sort	train_test_split
new / old (diff)	0.034889 / 0.037411 (-0.002522)	0.096096 / 0.014526 (0.081570)	0.111920 / 0.176557 (-0.064636)	0.166103 / 0.737135 (-0.571032)	0.111162 / 0.296338 (-0.185176)

Benchmark: benchmark_iterating.json

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
new / old (diff)	0.428037 / 0.215209 (0.212828)	4.294150 / 2.077655 (2.216495)	2.270331 / 1.504120 (0.766211)	2.108235 / 1.541195 (0.567041)	2.242560 / 1.468490 (0.774070)	0.489941 / 4.584777 (-4.094836)	3.688111 / 3.745712 (-0.057601)	3.450180 / 5.269862 (-1.819681)	2.175106 / 4.565676 (-2.390570)	0.057657 / 0.424275 (-0.366619)	0.007478 / 0.007607 (-0.000130)	0.505242 / 0.226044 (0.279198)	5.047817 / 2.268929 (2.778888)	2.724125 / 55.444624 (-52.720500)	2.419765 / 6.876477 (-4.456711)	2.723231 / 2.142072 (0.581159)	0.602382 / 4.805227 (-4.202846)	0.132362 / 6.500664 (-6.368302)	0.060600 / 0.075469 (-0.014869)

Benchmark: benchmark_map_filter.json

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
new / old (diff)	1.363356 / 1.841788 (-0.478431)	21.446474 / 8.074308 (13.372165)	15.074732 / 10.191392 (4.883340)	0.191837 / 0.680424 (-0.488587)	0.020565 / 0.534201 (-0.513636)	0.396692 / 0.579283 (-0.182591)	0.432390 / 0.434364 (-0.001974)	0.491747 / 0.540337 (-0.048591)	0.699203 / 1.386936 (-0.687733)