Datasets:

---

amannagrawall002

/

github-issues

like 0

[ "Thanks for reporting @TevenLeScao! I'm having a look...", "(not sure what just happened on the assignations sorry)", "For some reason this happens (both `datasets` version are on master) only on Python 3.6 and not Python 3.8.", "@TevenLeScao we are using `pandas` to serialize the dataset to JSON Lines. So it must be due to pandas. Could you please check the pandas version causing the issue?", "@TevenLeScao I have just checked it: this was a bug in `pandas` and it was fixed in version 1.2: https://github.com/pandas-dev/pandas/pull/36898", "Thanks ! I'm creating a PR", "Well I though it was me who has taken on this issue... 😅 ", "Sorry, I was also talking to teven offline so I already had the PR ready before noticing x)", "I was also already working in my PR... Nevermind. Next time we should pay attention if there is somebody (self-)assigned to an issue and if he/she is still working on it before overtaking it... 😄 ", "The fix is available on `master` @TevenLeScao , thanks for reporting" ]

[ "Updating to pyarrow-4.0.1 didn't fix the issue", "Here is an exemple dataset with 2 of these compressed JSON files: https://huggingface.co/datasets/thomwolf/github-python", "Hi @thomwolf, thanks for reporting.\r\n\r\nIt seems this might be due to the fact that the JSON Dataset builder uses `pyarrow.json` (`paj.read_json`) to read the data without using the Python standard `open(file,...` (which is the one patched with `xopen` to work in streaming mode).\r\n\r\nThis has to be fixed.", "Sorry for reopening this, but I'm having the same issue as @thomwolf when streaming a gzipped JSON Lines file from the hub. Or is that just not possible by definition?\r\nI installed `datasets`in editable mode from source (so probably includes the fix from #2608 ?): \r\n```\r\n>>> datasets.__version__\r\n'1.9.1.dev0'\r\n```\r\n\r\n```\r\n>>> msmarco = datasets.load_dataset(\"webis/msmarco\", \"corpus\", streaming=True)\r\nUsing custom data configuration corpus-174d3b7155eb68db\r\n>>> msmarco_iter = iter(msmarco['train'])\r\n>>> print(next(msmarco_iter))\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/media/ssd/TREC/msmarco/datasets/src/datasets/iterable_dataset.py\", line 338, in __iter__\r\n for key, example in self._iter():\r\n File \"/media/ssd/TREC/msmarco/datasets/src/datasets/iterable_dataset.py\", line 335, in _iter\r\n yield from ex_iterable\r\n File \"/media/ssd/TREC/msmarco/datasets/src/datasets/iterable_dataset.py\", line 78, in __iter__\r\n for key, example in self.generate_examples_fn(**self.kwargs):\r\n File \"/home/christopher/.cache/huggingface/modules/datasets_modules/datasets/msmarco/eb63dff8d83107168e973c7a655a6082d37e08d71b4ac39a0afada479c138745/msmarco.py\", line 96, in _generate_examples\r\n with gzip.open(file, \"rt\", encoding=\"utf-8\") as f:\r\n File \"/usr/lib/python3.6/gzip.py\", line 53, in open\r\n binary_file = GzipFile(filename, gz_mode, compresslevel)\r\n File \"/usr/lib/python3.6/gzip.py\", line 163, in __init__\r\n fileobj = self.myfileobj = builtins.open(filename, mode or 'rb')\r\nFileNotFoundError: [Errno 2] No such file or directory: 'https://huggingface.co/datasets/webis/msmarco/resolve/main/msmarco_doc_00.gz'\r\n```\r\n\r\nLoading the dataset without streaming set to True, works fine.", "Hi ! To make the streaming work, we extend `open` in the dataset builder to work with urls.\r\n\r\nTherefore you just need to use `open` before using `gzip.open`:\r\n```diff\r\n- with gzip.open(file, \"rt\", encoding=\"utf-8\") as f:\r\n+ with gzip.open(open(file, \"rb\"), \"rt\", encoding=\"utf-8\") as f:\r\n```\r\n\r\nYou can see that it is the case for oscar.py and c4.py for example:\r\n\r\nhttps://github.com/huggingface/datasets/blob/8814b393984c1c2e1800ba370de2a9f7c8644908/datasets/oscar/oscar.py#L358-L358\r\n\r\nhttps://github.com/huggingface/datasets/blob/8814b393984c1c2e1800ba370de2a9f7c8644908/datasets/c4/c4.py#L88-L88\r\n\r\n", "@lhoestq Sorry I missed that. Thank you Quentin!" ]

[ "#take" ]

[ "Hi !\r\nIf we want something more general, we could either\r\n1. delete the extracted files after the arrow data generation automatically, or \r\n2. delete each extracted file during the arrow generation right after it has been closed.\r\n\r\nSolution 2 is better to save disk space during the arrow generation. Is it what you had in mind ?\r\n\r\nThe API could look like\r\n```python\r\nload_dataset(..., delete_extracted_files_after_usage=True)\r\n```\r\n\r\nIn terms of implementation, here are some directions we could take for each solution:\r\n1. get the list of the extracted files from the DownloadManager and then delete them after the dataset is processed. This can be implemented in `download_and_prepare` I guess\r\n2. maybe wrap and mock `open` in the builder to make it delete the file when the file is closed.", "Also, if I delete the extracted files they need to be re-extracted again instead of loading from the Arrow cache files", "I think we already opened an issue about this topic (suggested by @stas00): duplicated of #2481?\r\n\r\nThis is in our TODO list... 😅 ", "I think the deletion of each extracted file could be implemented in our CacheManager and ExtractManager (once merged to master: #2295, #2277). 😉 ", "Oh yes sorry, I didn't check if this was a duplicate", "Nevermind @thomwolf, I just mentioned the other issue so that both appear linked in GitHub and we do not forget to close both once we make the corresponding Pull Request... That was the main reason! 😄 ", "Ok yes. I think this is an important feature to be able to use large datasets which are pretty much always compressed files.\r\n\r\nIn particular now this requires to keep the extracted file on the drive if you want to avoid reprocessing the dataset so in my case, this require using always ~400GB of drive instead of just 200GB (which is already significant). \r\n\r\nTwo nice features would be to:\r\n- allow to delete the extracted files without loosing the ability to load the dataset from the cached arrow-file\r\n- streamlined decompression when only the currently read file is extracted - this might require to read the list of files from the extracted archives before processing them?", "Here is a sample dataset with 2 such large compressed JSON files for debugging: https://huggingface.co/datasets/thomwolf/github-python", "Note that I'm confirming that with the current master branch of dataset, deleting extracted files (without deleting the arrow cache file) lead to **re-extracting** these files when reloading the dataset instead of directly loading the arrow cache file.", "Hi ! That's weird, it doesn't do that on my side (tested on master on my laptop by deleting the `extracted` folder in the download cache directory). You tested with one of the files at https://huggingface.co/datasets/thomwolf/github-python that you have locally ?", "Yes it’s when I load local compressed JSON line files with load_dataset(‘json’, data_files=…) ", "@thomwolf I'm sorry but I can't reproduce this problem. I'm also using: \r\n```python\r\nds = load_dataset(\"json\", split=\"train\", data_files=data_files, cache_dir=cache_dir)\r\n```\r\nafter having removed the extracted files:\r\n```python\r\nassert sorted((cache_dir / \"downloads\" / \"extracted\").iterdir()) == []\r\n```\r\n\r\nI get the logging message:\r\n```shell\r\nWARNING datasets.builder:builder.py:531 Reusing dataset json ...\r\n```", "Do you confirm the extracted folder stays empty after reloading?", "> \r\n> \r\n> Do you confirm the extracted folder stays empty after reloading?\r\n\r\nYes, I have the above mentioned assertion on the emptiness of the extracted folder:\r\n```python\r\nassert sorted((cache_dir / \"downloads\" / \"extracted\").iterdir()) == []\r\n```\r\n" ]

[]

[ "Hi @erikadistefano , thanks for reporting the issue.\r\n\r\nI have created a Pull Request that should fix it. \r\n\r\nOnce merged into master, feel free to update your installed `datasets` library (either by installing it from our GitHub master branch or waiting until our next release) to be able to load omp dataset." ]

[ "Hi ! Do you have an example in mind that shows how this could be useful ?", "Example:\n\nMerge 2 datasets into one datasets\n\nLabel extraction from dataset\n\ndataset(text, label)\n —> dataset(text, newlabel)\n\nTextCleaning.\n\n\nFor image dataset, \nTransformation are easier (ie linear algebra).\n\n\n\n\n\n\n> On Jul 6, 2021, at 17:39, Quentin Lhoest ***@***.***> wrote:\n> \n> \n> Hi ! Do you have an example in mind that shows how this could be useful ?\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "There are already a few transformations that you can apply on a dataset using methods like `dataset.map()`.\r\nYou can find examples in the documentation here:\r\nhttps://huggingface.co/docs/datasets/processing.html\r\n\r\nYou can merge two datasets with `concatenate_datasets()` or do label extraction with `dataset.map()` for example", "Ok, sure.\n\nThanks for pointing on functional part.\nMy question is more\n“Philosophical”/Design perspective.\n\nThere are 2 perspetive:\n Add transformation methods to \n Dataset Class\n\n\n OR Create a Transformer Class\n which operates on Dataset Class.\n\nT(Dataset) —> Dataset\n\ndatasetnew = MyTransform.transform(dataset)\ndatasetNew.save(path)\n\n\nWhat would be the difficulty\nof implementing a Transformer Class\noperating at dataset level ?\n\n\nthanks\n\n\n\n\n\n\n\n\n\n> On Jul 6, 2021, at 22:00, Quentin Lhoest ***@***.***> wrote:\n> \n> \n> There are already a few transformations that you can apply on a dataset using methods like dataset.map().\n> You can find examples in the documentation here:\n> https://huggingface.co/docs/datasets/processing.html\n> \n> You can merge two datasets with concatenate_datasets() or do label extraction with dataset.map() for example\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "I can imagine that this would be a useful API to implement processing pipelines as transforms. They could be used to perform higher level transforms compared to the atomic transforms allowed by methods like map, filter, etc.\r\n\r\nI guess if you find any transform that could be useful for text dataset processing, image dataset processing etc. we could definitely start having such transforms :)", "Thanks for reply.\n\nWhat would be the constraints\nto have\nDataset —> Dataset consistency ?\n\nMain issue would be\nlarger than memory dataset and\nserialization on disk.\n\nTechnically,\none still process at atomic level\nand try to wrap the full results\ninto Dataset…. (!)\n\nWhat would you think ?\n\n\n\n\n\n\n\n\n> On Jul 7, 2021, at 16:51, Quentin Lhoest ***@***.***> wrote:\n> \n> \n> I can imagine that this would be a useful API to implement processing pipelines as transforms. They could be used to perform higher level transforms compared to the atomic transforms allowed by methods like map, filter, etc.\n> \n> I guess if you find any transform that could be useful for text dataset processing, image dataset processing etc. we could definitely start having such transforms :)\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "We can be pretty flexible and not impose any constraints for transforms.\r\n\r\nMoreover, this library is designed to support datasets bigger than memory. The datasets are loaded from the disk via memory mapping, without filling up RAM. Even processing functions like `map` work in a batched fashion to not fill up your RAM. So this shouldn't be an issue", "Ok thanks.\n\nBut, Dataset has various flavors.\nIn current design of Dataset,\n how the serialization on disk is done (?)\n\n\nThe main issue is serialization \nof newdataset= Transform(Dataset)\n (ie thats why am referring to Out Of memory dataset…):\n\n Should be part of Transform or part of dataset ?\n\n\n\n\nMaybe, not, since the output is aimed to feed model in memory (?)\n\n\n\n\n\n\n\n\n> On Jul 7, 2021, at 18:04, Quentin Lhoest ***@***.***> wrote:\n> \n> \n> We can be pretty flexible and not impose any constraints for transforms.\n> \n> Moreover, this library is designed to support datasets bigger than memory. The datasets are loaded from the disk via memory mapping, without filling up RAM. Even processing functions like map work in a batched fashion to not fill up your RAM. So this shouldn't be an issue\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "I'm not sure I understand, could you elaborate a bit more please ?\r\n\r\nEach dataset is a wrapper of a PyArrow Table that contains all the data. The table is loaded from an arrow file on the disk.\r\nWe have an ArrowWriter and ArrowReader class to write/read arrow tables on disk or in in-memory buffers." ]

[ "Hi @profsatwinder.\r\n\r\nIt looks like you are using an old version of `datasets`. Please update it with `pip install -U datasets` and indicate if the problem persists.", "@albertvillanova Thanks for the information. I updated it to 1.9.0 and the issue is resolved. Thanks again. " ]

[ "Hi! I'm transferring this issue over to `datasets`", "I'm using the datasets concatenate dataset to combine the datasets and then train.\r\ntrain_dataset = concatenate_datasets([dataset1, dataset2, common_voice_train])\r\n\r\n", "Hi @BirgerMoell.\r\n\r\nYou have several options:\r\n- to set caching to be stored on a different path location, other than the default one (`~/.cache/huggingface/datasets`):\r\n - either setting the environment variable `HF_DATASETS_CACHE` with the path to the new cache location\r\n - or by passing it with the parameter `cache_dir` when loading each of the datasets: `dataset = load_dataset(..., cache_dir=your_new_location)`\r\n\r\n You can get all the information in the docs: https://huggingface.co/docs/datasets/loading_datasets.html#cache-directory\r\n- I wouldn't recommend disabling caching, because current implementation generates cache files anyway, although in a temporary directory and they are deleted when the session closes. See details here: https://huggingface.co/docs/datasets/processing.html#enable-or-disable-caching\r\n- You could alternatively load the datasets in streaming mode. This is a new feature which allows loading the datasets without downloading the entire files. More information here: https://huggingface.co/docs/datasets/dataset_streaming.html", "Hi @BirgerMoell,\r\n\r\nWe are planning to add a new feature to datasets, which could be interesting in your case: Add the option to delete temporary files (decompressed files) from the cache directory (see: #2481, #2604).\r\n\r\nWe will ping you once this feature is implemented, so that the size of your cache directory will be considerably reduced." ]

[ "Hi @mmajurski, thanks for reporting this issue.\r\n\r\nIndeed this misalignment arises because the source dataset context field contains leading blank spaces (and these are counted within the answer_start), while our datasets loading script removes these leading blank spaces.\r\n\r\nI'm going to fix our script so that all leading blank spaces in the source dataset are kept, and there is no misalignment between the answer text and the answer_start within the context.", "If you are going to be altering the data cleaning from the source Squad dataset, here is one thing to consider.\r\nThere are occasional double spaces separating words which it might be nice to get rid of. \r\n\r\nEither way, thank you." ]

[ "Hi ! This is because you first need to format the dataset for pytorch:\r\n\r\n```python\r\n>>> import torch\r\n>>> from datasets import load_dataset\r\n>>> dataset = load_dataset('oscar', \"unshuffled_deduplicated_en\", split='train', streaming=True)\r\n>>> torch_iterable_dataset = dataset.with_format(\"torch\")\r\n>>> assert isinstance(torch_iterable_dataset, torch.utils.data.IterableDataset)\r\n>>> dataloader = torch.utils.data.DataLoader(torch_iterable_dataset, batch_size=4)\r\n>>> next(iter(dataloader))\r\n{'id': tensor([0, 1, 2, 3]), 'text': ['Mtendere Village was inspired...]}\r\n```\r\n\r\nThis is because the pytorch dataloader expects a subclass of `torch.utils.data.IterableDataset`. Since you can't pass an arbitrary iterable to a pytorch dataloader, you first need to build an object that inherits from `torch.utils.data.IterableDataset` using `with_format(\"torch\")` for example.\r\n", "Thank you for that and the example! \r\n\r\nWhat you said makes total sense; I just somehow missed that and assumed HF IterableDataset was a subclass of Torch IterableDataset. " ]

[ "This was actually a second error arising from a too small block-size in the json reader.\r\n\r\nFinding the right block size is difficult for the layman user" ]

[ "I am trying to load a dataset using Hugging Face Datasets load_dataset method. I am getting the value error as show below. Can someone help with this? I am using Windows laptop and Google Colab notebook.\r\n\r\n```\r\n!pip install zstandard\r\nfrom datasets import load_dataset\r\n\r\nlds = load_dataset(\r\n \"json\",\r\n data_files=\"https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst\",\r\n split=\"train\",\r\n streaming=True,\r\n)\r\n\r\nWARNING:datasets.builder:Using custom data configuration default-a1d9e8eaedd958cd\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n[<ipython-input-12-5b4fdcb8e6d5>](https://localhost:8080/#) in <module>\r\n 6 )\r\n 7 \r\n----> 8 next(iter(law_dataset_streamed))\r\n\r\n17 frames\r\n[/usr/local/lib/python3.8/dist-packages/fsspec/core.py](https://localhost:8080/#) in get_compression(urlpath, compression)\r\n 485 compression = infer_compression(urlpath)\r\n 486 if compression is not None and compression not in compr:\r\n--> 487 raise ValueError(\"Compression type %s not supported\" % compression)\r\n 488 return compression\r\n 489 \r\n\r\nValueError: Compression type zstd not supported\r\n```", "I just tried on google colab and this works:\r\n```python\r\n!pip install zstandard\r\n!pip install datasets\r\nfrom datasets import load_dataset\r\n\r\nlds = load_dataset(\r\n \"json\",\r\n data_files=\"https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst\",\r\n split=\"train\",\r\n streaming=True,\r\n)\r\nnext(iter(lds))\r\n```\r\n\r\nCan you check that you have a correct installation of `zstandard` ?", "@lhoestq please note [this](https://github.com/huggingface/datasets/issues/2572#issuecomment-1363718916) is a duplicate of:\r\n- #5388", "Oh thanks I missed that one !", "> I just tried on google colab and this works:\r\n> \r\n> ```python\r\n> !pip install zstandard\r\n> !pip install datasets\r\n> from datasets import load_dataset\r\n> \r\n> lds = load_dataset(\r\n> \"json\",\r\n> data_files=\"https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst\",\r\n> split=\"train\",\r\n> streaming=True,\r\n> )\r\n> next(iter(lds))\r\n> ```\r\n> \r\n> Can you check that you have a correct installation of `zstandard` ?\r\n\r\nI was downloading datasets first then was doing zstandard installation and that was causing the issue. This was highlighted by the Hugging Face staff and that helped. Now the issue is resolved. Thank you." ]

[ "Hi @suzyahyah, thanks for reporting.\r\n\r\nThe message you get is indeed not an error message, but a warning coming from Hugging Face `transformers`. The complete warning message is:\r\n```\r\nSome weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.dense.weight', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.bias', 'lm_head.layer_norm.weight']\r\n- This IS expected if you are initializing RobertaModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing RobertaModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\n```\r\n\r\nIn this case, this behavior IS expected and you can safely ignore the warning message.\r\n\r\nThe reason is that you are just using RoBERTa to get the contextual embeddings of the input sentences/tokens, thus leaving away its head layer, whose weights are ignored.\r\n\r\nFeel free to reopen this issue if you need further explanations.", "Hi @suzyahyah, I have created a Pull Request to filter out that warning message in this specific case, since the behavior is as expected and the warning message can only cause confusion for users (as in your case)." ]

[ "It is probably worth noting here that the [documentation](https://huggingface.co/docs/datasets/process#concatenate) is misleading (indicating that it does work for IterableDatasets):\r\n\r\n> You can also mix several datasets together by taking alternating examples from each one to create a new dataset. This is known as interleaving, and you can use it with [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets). **Both [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets) and [concatenate_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.concatenate_datasets) will work with regular [Dataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.Dataset) and [IterableDataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.IterableDataset) objects**. Refer to the [Stream](https://huggingface.co/docs/datasets/stream#interleave) section for an example of how it’s used. ", "Thanks for the heads up, I'll fix that" ]

[]

[ "Hi ! I just tried to reproduce what you said:\r\n- create a local builder class\r\n- use `load_dataset`\r\n- update the builder class code\r\n- use `load_dataset` again (with or without `ignore_verifications=True`)\r\nAnd it creates a new cache, as expected.\r\n\r\nWhat modifications did you do to your builder's code ?", "Hi @lhoestq. Thanks for your reply. I just did minor modifications for which it should not regenerate cache (for e.g. Adding a print statement). Overall, regardless of cache miss, there should be an explicit option to allow reuse of existing cache if author knows cache shouldn't be affected.", "The cache is based on the hash of the dataset builder's code, so changing the code makes it recompute the cache.\r\n\r\nYou could still rename the cache directory of your previous computation to the new expected cache directory if you want to avoid having to recompute it and if you're sure that it would generate the exact same result.\r\n\r\nThe verifications are data integrity verifications: it checks the checksums of the downloaded files, as well as the size of the generated splits.", "Hi @apsdehal,\r\n\r\nIf you decide to follow @lhoestq's suggestion to rename the cache directory of your previous computation to the new expected cache directory, you can do the following to get the name of the new expected cache directory once #2500 is merged:\r\n```python\r\nfrom datasets import load_dataset_builder\r\ndataset_builder = load_dataset_builder(\"path/to/your/dataset\")\r\nprint(dataset_builder.cache_dir)\r\n```\r\n\r\nThis way, you don't have to recompute the hash of the dataset script yourself each time you modify the script." ]

[ "Hi ! Can you share the function you pass to `map` ?\r\nI know you mentioned it would be hard to share some code but this would really help to understand what happened", "This is the same behavior as in #4883, so I'm closing this issue as a duplicate. " ]

[ "excuse me, my `datasets` version is `2.2.2`, but I also just see the error info like \r\n```\r\nDuplicatedKeysError: FAILURE TO GENERATE DATASET !\r\nFound duplicate Key: 0\r\nKeys should be unique and deterministic in nature\r\n```", "Hi ! for which dataset do you have this error ?\r\n\r\nAlso note that this issue is just about improving the error message, which is not very friendly x)", "@lhoestq I would like to take a hit at improving the error message. Will open a draft PR and will reach out to you for review\r\n", "> DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.\r\n\r\n@lhoestq when you mention 42th and 1337th in the above case , are these values the examples' \"id\" or are they the examples' index ? ", "Hi ! Thanks @VijayKalmath :)\r\n\r\nIn the general case, examples don't have an \"id\" field, so I think it should correspond to the index", "@lhoestq , I have opened a draft PR for this Issue. \r\n\r\nI wanted to check with you if there is a way to get `<path/to/the/dataset/script>` currently or do I need to add extra code to find that. \r\n\r\nIf I need to find the script , I can assume that the generator function will always be in `datasets/{dataset_name}/{dataset_name}.py`. ", "Thanks !\r\n\r\n> I wanted to check with you if there is a way to get <path/to/the/dataset/script> currently or do I need to add extra code to find that.\r\n\r\nYou don't have access to this info inside the ArrowWriter unfortunately. This info is available in builder.py in the DatasetBuilder code that uses the ArrowWriter though, maybe a try-catch there can do the job" ]

[ "Hi @GuillemGSubies, thanks for reporting.\r\n\r\nI have made a PR to fix this issue and allow metrics to be computed also for multilabel classification problems.", "Looks nice, thank you very much! 🚀 ", "Sorry for reopening but I just noticed that the `_compute` method for the F1 metric is still not good enough for multilabel problems:\r\n\r\nhttps://github.com/huggingface/datasets/blob/92a3ee549705aa0a107c9fa5caf463b3b3da2616/metrics/f1/f1.py#L115\r\n\r\nSomehow we should be able to change the parameter `average` at least", "@GuillemGSubies, the parameter `average` passed to `_compute` is then passed to `f1_score`. This is right." ]

[ "Hi ! Thanks for reporting. This is due to the WebNLG repository that got updated today.\r\nI just pushed a fix at #2558 - this shouldn't happen anymore in the future.", "This is fixed on `master` now :)\r\nWe'll do a new release soon !" ]

[ "Two questions:\r\n- with `datasets-cli env` we don't have any information on the dataset script version used. Should we give access to this somehow? Either as a note in the Error message or as an argument with the name of the dataset to `datasets-cli env`?\r\n- I don't really understand why the id is duplicated in the code of `code_search_net`, how can I debug this actually?", "Thanks for reporting. There was indeed an issue with the keys. The key was the addition of the file id and row id, which resulted in collisions. I just opened a PR to fix this at https://github.com/huggingface/datasets/pull/2555\r\n\r\nTo help users debug this kind of errors we could try to show a message like this\r\n```python\r\nDuplicateKeysError: both 42th and 1337th examples have the same keys `48`.\r\nPlease fix the dataset script at <path/to/the/dataset/script>\r\n```\r\n\r\nThis way users who what to look for if they want to debug this issue. I opened an issue to track this: https://github.com/huggingface/datasets/issues/2556", "and are we sure there are not a lot of datasets which are now broken with this change?", "Thanks to the dummy data, we know for sure that most of them work as expected.\r\n`code_search_net` wasn't caught because the dummy data only have one dummy data file while the dataset script can actually load several of them using `os.listdir`. Let me take a look at all the other datasets that use `os.listdir` to see if the keys are alright", "I found one issue on `fever` (PR here: https://github.com/huggingface/datasets/pull/2557)\r\nAll the other ones seem fine :)", "Hi! Got same error when loading other dataset:\r\n```python3\r\nload_dataset('wikicorpus', 'raw_en')\r\n```\r\n\r\ntb:\r\n```pytb\r\n---------------------------------------------------------------------------\r\nDuplicatedKeysError Traceback (most recent call last)\r\n/opt/conda/lib/python3.8/site-packages/datasets/builder.py in _prepare_split(self, split_generator)\r\n 1109 example = self.info.features.encode_example(record)\r\n-> 1110 writer.write(example, key)\r\n 1111 finally:\r\n\r\n/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in write(self, example, key, writer_batch_size)\r\n 341 if self._check_duplicates:\r\n--> 342 self.check_duplicate_keys()\r\n 343 # Re-intializing to empty list for next batch\r\n\r\n/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in check_duplicate_keys(self)\r\n 352 if hash in tmp_record:\r\n--> 353 raise DuplicatedKeysError(key)\r\n 354 else:\r\n\r\nDuplicatedKeysError: FAILURE TO GENERATE DATASET !\r\nFound duplicate Key: 519\r\nKeys should be unique and deterministic in nature\r\n```\r\n\r\nVersion: datasets==1.11.0", "Fixed by #2555.", "The wikicorpus issue has been fixed by https://github.com/huggingface/datasets/pull/2844\r\n\r\nWe'll do a new release of `datasets` soon :)" ]

[]

[ "Hi @nelson-liu,\r\n\r\nYou can pass the parameter `features` to `load_dataset`: https://huggingface.co/docs/datasets/_modules/datasets/load.html#load_dataset\r\n\r\nIf you look at the code of the MNLI script you referred in your question (https://github.com/huggingface/datasets/blob/master/datasets/multi_nli/multi_nli.py#L62-L77), you can see how the Features were originally specified. \r\n\r\nFeel free to use it as a template, customize it and pass it to `load_dataset` using the parameter `features`.", "ah got it, thanks!" ]

[ "Hi @luyug, thanks for reporting.\r\n\r\nThe good news is that we fixed this issue only 9 days ago: #2507.\r\n\r\nThe patch is already in the master branch of our repository and it will be included in our next `datasets` release version 1.9.0.\r\n\r\nFeel free to reopen the issue if the problem persists." ]

[ "Hi ! It looks like an issue with the virtual disk you are using.\r\n\r\nWe load datasets using memory mapping. In general it makes it possible to load very big files instantaneously since it doesn't have to read the file (it just assigns virtual memory to the file on disk).\r\nHowever there happens to be issues with virtual disks (for example on spot instances), for which memory mapping does a pass over the entire file, and this takes a while. We are discussing about this issue here: #2252 \r\n\r\nMemory mapping is something handled by the OS so we can't do much about it, though we're still trying to figure out what's causing this behavior exactly to see what we can do.", "Okay, that's exactly my case, with spot instances... Therefore this isn't something we can change in any way to be able to load the dataset faster? I mean, what do you do internally at huggingface for being able to use spot instances with datasets efficiently?", "There are no solutions yet unfortunately.\r\nWe're still trying to figure out a way to make the loading instantaneous on such disks, I'll keep you posted" ]

[ "Hi @stas00, thanks for pointing out this issue with logging.\r\n\r\nI agree that `datasets` can sometimes be too verbose... I can create a PR and we could discuss there the choice of the log levels for different parts of the code." ]

[ "very much related: https://github.com/huggingface/datasets/pull/2333", "Hi @VictorSanh, thank you for reporting this issue with duplicated keys.\r\n\r\n- The issue with \"adversarial_qa\" was fixed 23 days ago: #2433. Current version of `datasets` (1.8.0) includes the patch.\r\n- I am investigating the issue with `drop`. I'll ping you to keep you informed.", "Hi @VictorSanh, the issue is already fixed and merged into master branch and will be included in our next release version 1.9.0.", "thank you!" ]

[ "Hi ! `load_dataset` downloads the full dataset once and caches it, so that subsequent calls to `load_dataset` just reloads the dataset from your disk.\r\nThen when you specify a `split` in `load_dataset`, it will just load the requested split from the disk. If your specified split is a sliced split (e.g. `\"train[:10]\"`), then it will load the 10 first rows of the train split that you have on disk.\r\n\r\nTherefore, as long as you don't delete your cache, all your calls to `load_dataset` will be very fast. Except the first call that downloads the dataset of course ^^", "That’s a use case for the new streaming feature, no?", "Hi @reachtarunhere.\r\n\r\nBesides the above insights provided by @lhoestq and @thomwolf, there is also a Dataset feature in progress (I plan to finish it this week): #2249, which will allow you, when calling `load_dataset`, to pass the option to download/preprocess/cache only some specific split(s), which will definitely speed up your workflow.\r\n\r\nIf this feature is interesting for you, I can ping you once it will be merged into the master branch.", "Thanks all for responding.\r\n\r\nHey @albertvillanova \r\n\r\nThanks. Yes, I would be interested.\r\n\r\n@lhoestq I think even if a small split is specified it loads up the full dataset from the disk (please correct me if this is not the case). Because it does seem to be slow to me even on subsequent calls. There is no repeated downloading so it seems that the cache is working.\r\n\r\nI am not aware of the streaming feature @thomwolf mentioned. So I might need to read up on it.", "@reshinthadithyan I use the .select function to have a fraction of indices.", "If I want to create a dataset, containing only the 10 elements of a given dataset (slice it), how do I do that?", "```python \r\nsmall_ds = ds.select(range(10))\r\n```", "\r\n\r\n> ```python\r\n> small_ds = ds.select(range(10))\r\n> ```\r\n\r\nThanks, but this doesn't help me to save time during initial loading, right?", "Indeed by default load_dataset would download and prepare everything as Arrow files. And passing `split=train[:10]` memory maps only the beginning of the full dataset that has been prepared on disk.\r\n\r\nIf you don't want to download everything, you can use streaming : \r\n```python \r\nids = load_dataset(..., streaming=True)\r\nfirst_samples = list(ids[\"train\"].take(10))\r\n```\r\n\r\nTo get a Dataset you can use \r\n```python \r\nds = Dataset.from_generator(ids.take(10).__iter__)\r\n```\r\n\r\nedit: fixed small bug", "Thanks @lhoestq, but I don't think it is 100% accurate, as it doesn't keep the dataset structure exactly the same.\r\nTo load the full dataset, I do:\r\n```\r\ndata = load_dataset(\"json\", data_files=\"a.json\")\r\ntrain_data = data[\"train\"].shuffle()\r\n```\r\n\r\nBut when I am changing it as per your instructions: \r\n```\r\nids = load_dataset(\"json\", data_files=\"a.json\", streaming=True)\r\ndata = Dataset.from_generator(ids[\"train\"].take(1).__iter__)\r\ntrain_data = data[\"train\"].shuffle()\r\n```\r\nIt throws KeyError.\r\nI need a simple way, like you suggested, to have a subset of a Dataset, which exactly the same attributes.\r\n", "Whoops I fixed my code sorry\r\n```diff\r\n- ds = Dataset.from_generator(ids[\"train\"].take(10).__iter__)\r\n+ ds = Dataset.from_generator(ids.take(10).__iter__)\r\n```\r\n\r\nin your case that means running\r\n```python\r\ntrain_data = data.shuffle()\r\n```\r\n\r\nwithout `[\"train\"]`" ]

[ "I'm just retaking and working on #2324. 😉 ", "Resolved via https://github.com/huggingface/datasets/pull/4006." ]

[ "Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**?", "> Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**?\r\n\r\nOh, I am sorry\r\nI would reopen the post on huggingface/transformers" ]

[ "Hi @joshzwiebel, thanks for reporting. We are going to align with `transformers`." ]

[ "I'm currently adding it, the entire dataset is quite big around 30 GB so I add splits separately. You can take a look here https://huggingface.co/datasets/merve/coco", "I talked to @lhoestq and it's best if I download this dataset through TensorFlow datasets instead, so I'll be implementing that one really soon.\r\n@NielsRogge ", "I started adding COCO, will be done tomorrow EOD\r\nmy work so far https://github.com/merveenoyan/datasets (my fork)", "Hi Merve @merveenoyan , thank you so much for your great contribution! May I ask about the current progress of your implementation? Cuz I see the pull request is still in progess here. Or can I just run the COCO scripts in your fork repo?", "Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon. ", "> Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon.\r\n\r\nIt's really nice of you!! I see you've commited another version just now", "@yixuanren we're working on it, will be available soon, thanks a lot for your patience", "Hi @NielsRogge and @merveenoyan, did you find a way to load a dataset with COCO annotations to HF's hub?\r\nI have a panoptic segmentation dataset in COCO format and would like to share it with the community.\r\nThanks in advance :)", "The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script)", "> The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script)\r\n\r\nHi @lhoestq , thank you for your quick reply.\r\nI've correctly created a metadata.jsonl file for a dataset with instance segmentation annotations [here](https://huggingface.co/datasets/lombardata/data_2017)\r\nbut do not understand how I can integrate panoptic annotations with the metadata format of ImageFolder datasets. The \"problem\" with panoptic annotations is that we have a folder with images, a json file with annotations and another folder with png annotations.\r\n\r\nI checked between all the datasets already published on HuggingFace and, the only one who has uploaded a correct panoptic dataset is @NielsRogge [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo). Indeed he accomplished to have three fields : \r\n1.image (image)\r\n2.label (image)\r\n3.segments_info (list)\r\nbut I not find the corresponding code that allows to upload a panoptic dataset from this 3 sources.\r\nCan you please share an example code?\r\nThanks !", "Both were uploaded using `ds.push_to_hub()` :)\r\n\r\nYou can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column(\"image\", Image())`.\r\n\r\n```python\r\nfrom datasets import Dataset, Image\r\n\r\nds = Dataset.from_dict(...)\r\nds = ds.cast_column(\"image\", Image())\r\nds = ds.cast_column(\"label\", Image())\r\nds.push_to_hub(...)\r\n```", "> Both were uploaded using `ds.push_to_hub()` :)\r\n> \r\n> You can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column(\"image\", Image())`.\r\n> \r\n> ```python\r\n> from datasets import Dataset, Image\r\n> \r\n> ds = Dataset.from_dict(...)\r\n> ds = ds.cast_column(\"image\", Image())\r\n> ds = ds.cast_column(\"label\", Image())\r\n> ds.push_to_hub(...)\r\n> ```\r\n\r\nThank you very much @lhoestq , I succesfully created a hf dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_21) with the two fields :\r\n1.image (image)\r\n2.label (image)\r\nfollowing your suggestions. Now still remain the problem of uploading **segments_info** information to the dataset.\r\nThere is a function that easily imports the _panoptic_coco_annotation.json_ file to a segment_info field?\r\nI think we must define a **list_of_segment**, i.e. a list of lists of this type : \r\n```python\r\n[ { \"area\": 214858, \"bbox\": [ 0, 0, 511, 760 ], \"category_id\": 0, \"id\": 7895160, \"iscrowd\": 0 }, { \"area\": 73067, \"bbox\": [ 98, 719, 413, 253 ], \"category_id\": 3, \"id\": 3289680, \"iscrowd\": 0 }, { \"area\": 832, \"bbox\": [ 53, 0, 101, 16 ], \"category_id\": 5, \"id\": 5273720, \"iscrowd\": 0 }, { \"area\": 70668, \"bbox\": [ 318, 60, 191, 392 ], \"category_id\": 8, \"id\": 15132390, \"iscrowd\": 0 }, { \"area\": 32696, \"bbox\": [ 0, 100, 78, 872 ], \"category_id\": 18, \"id\": 472063, \"iscrowd\": 0 }, { \"area\": 76045, \"bbox\": [ 42, 48, 264, 924 ], \"category_id\": 37, \"id\": 16713830, \"iscrowd\": 0 }, { \"area\": 27103, \"bbox\": [ 288, 482, 216, 306 ], \"category_id\": 47, \"id\": 16753408, \"iscrowd\": 0 } ]\r\n```\r\nand then apply again the **cast_column** function [here](https://github.com/huggingface/datasets/blob/2.13.0/src/datasets/arrow_dataset.py#L2060) but with a list as a second argument, like : \r\n```python\r\nfrom datasets import Dataset, Image\r\nds = ds.cast_column(\"image\", Image())\r\nds = ds.cast_column(\"label\", Image())\r\nds = ds.cast_column(\"segments_info\", list)\r\n```\r\nbut I do not see how to transfer the information of the _panoptic_coco_annotation.json_ to a list of lists of this type : \r\n```python\r\n[ { \"area\": 214858, \"bbox\": [ 0, 0, 511, 760 ], \"category_id\": 0, \"id\": 7895160, \"iscrowd\": 0 }, { \"area\": 73067, \"bbox\": [ 98, 719, 413, 253 ], \"category_id\": 3, \"id\": 3289680, \"iscrowd\": 0 }, { \"area\": 832, \"bbox\": [ 53, 0, 101, 16 ], \"category_id\": 5, \"id\": 5273720, \"iscrowd\": 0 }, { \"area\": 70668, \"bbox\": [ 318, 60, 191, 392 ], \"category_id\": 8, \"id\": 15132390, \"iscrowd\": 0 }, { \"area\": 32696, \"bbox\": [ 0, 100, 78, 872 ], \"category_id\": 18, \"id\": 472063, \"iscrowd\": 0 }, { \"area\": 76045, \"bbox\": [ 42, 48, 264, 924 ], \"category_id\": 37, \"id\": 16713830, \"iscrowd\": 0 }, { \"area\": 27103, \"bbox\": [ 288, 482, 216, 306 ], \"category_id\": 47, \"id\": 16753408, \"iscrowd\": 0 } ]\r\n```\r\nlike @NielsRogge has done [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo).\r\nThank you again for your help and have a good day !", "You can pass this data in .from_dict() - no need to cast anything for this column\r\n\r\n```python\r\nds = Dataset.from_dict({\r\n \"image\": [...],\r\n \"label\": [...],\r\n \"segments_info\": [...],\r\n)}\r\n```\r\n\r\nwhere `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts.", "> You can pass this data in .from_dict() - no need to cast anything for this column\r\n> \r\n> ```python\r\n> ds = Dataset.from_dict({\r\n> \"image\": [...],\r\n> \"label\": [...],\r\n> \"segments_info\": [...],\r\n> )}\r\n> ```\r\n> \r\n> where `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts.\r\n\r\nThank you for the quick reply @lhoestq , but then how to generate the `segments_info` list of lists of dicts starting from a _panoptic_coco_annotation.json_ file ?\r\n\r\n\r\n", "You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere", "> You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere\r\n\r\nPerfect, I've done it and succesfully uploaded a new dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_22), but I've (I hope) a last problem.\r\nThe dataset has currently 302 images and, when I upload it to the hub, only the first page of images is correctly uploaded.\r\nWhen I try to see the second/third/fourth page of items of my dataset, I can see that the fields **segments_info** and **image_name** are correctly uploaded, while the images are not (the \"null\" string is shown everywhere).\r\n\r\nI've checked the path of images that are not uploaded and they exists, is there a problem with the size of the dataset ?\r\nHow can I upload the whole dataset to the hub ?\r\nThank you again @lhoestq and have a good day !", "Awesome ! Your dataset looks all good 🤗 \r\n\r\nThe `null` in the viewer is a bug on our side, let me investigate" ]

[]

[ "Hi,\r\n\r\nthis issue has been already reported in the dataset repo (https://github.com/dair-ai/emotion_dataset/issues/2), so this is a bug on their side.", "The documentation has another bug in the dataset card [here](https://huggingface.co/datasets/emotion). \r\n\r\nIn the dataset summary **six** emotions are mentioned: *\"six basic emotions: anger, fear, joy, love, sadness, and surprise\"*, however, in the datafields section we have only **five**:\r\n```\r\nlabel: a classification label, with possible values including sadness (0), joy (1), love (2), anger (3), fear (4).\r\n```", "@GDGauravDutta the dataset author replied in their issue: https://github.com/dair-ai/emotion_dataset/issues/2\r\n> The dataset released is a preprocessed six emotion variant released for educational and research purposes.\r\n\r\n@albertovilla the dataset card was fixed with 6 emotions." ]

[ "The `task_templates` API is deprecated in favor of the `train-eval-index` YAML field, so I'm closing this issue." ]

[ "Hi ! `map` calls `__getstate__` using `dill` to hash your map function. This is used by the caching mechanism to recover previously computed results. That's why you don't see any `__setstate__` call.\r\n\r\nWhy do you change an attribute of your tokenizer when `__getstate__` is called ?", "@lhoestq because if I try to pickle my custom tokenizer (it contains a pure python pretokenization step in an otherwise rust backed tokenizer) I get\r\n\r\n> Exception: Error while attempting to pickle Tokenizer: Custom PreTokenizer cannot be serialized\r\n\r\nSo I remove the Custom PreTokenizer in `__getstate__` and then restore it in `__setstate__` (since it doesn't contain any state). This is what my `__getstate__` / `__setstate__` looks like:\r\n\r\n def __getstate__(self):\r\n \"\"\"\r\n Removes pre_tokenizer since it cannot be pickled\r\n \"\"\"\r\n logger.debug(\"Copy state dict\")\r\n out = self.__dict__.copy()\r\n logger.debug(\"Detaching pre_tokenizer\")\r\n out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([]) \r\n return out\r\n\r\n def __setstate__(self, d):\r\n \"\"\"\r\n Reinstates pre_tokenizer\r\n \"\"\"\r\n logger.debug(\"Reattaching pre_tokenizer\")\r\n self.__dict__ = d\r\n self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()\r\n\r\nIf this is the case can you think of another way of avoiding my issue?", "Actually, maybe I need to deep copy `self.__dict__`? That way `self` isn't modified. That was my intention and I thought it was working - I'll double-check after the weekend.", "Doing a deep copy results in the warning:\r\n\r\n> 06/20/2021 16:02:15 - WARNING - datasets.fingerprint - Parameter 'function'=<function tokenize_function at 0x7f1e95f05d40> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.\r\n\r\n\r\n```\r\ndef __getstate__(self):\r\n \"\"\"\r\n Removes pre_tokenizer since it cannot be pickled\r\n \"\"\"\r\n logger.debug(\"Copy state dict\")\r\n out = copy.deepcopy(self.__dict__)\r\n logger.debug(\"Detaching pre_tokenizer\")\r\n out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([]) \r\n return out\r\n```", "Looks like there is still an object that is not pickable in your `tokenize_function` function.\r\n\r\nYou can test if an object can be pickled and hashed by using \r\n```python\r\nfrom datasets.fingerprint import Hasher\r\n\r\nHasher.hash(my_object)\r\n```\r\n\r\nUnder the hood it pickles the object to compute its hash, so it calls `__getstate__` when applicable.", "I figured it out, the problem is deep copy itself uses pickle (unless you implement `__deepcopy__`). So when I changed `__getstate__` it started throwing an error.\r\n\r\nI'm sure there's a better way of doing this, but in order to return the `__dict__` without the non-pikelable pre-tokeniser and without modifying self I removed the pre-tokenizers, did a deep copy and then re-generated it.\r\n\r\nIt does work - although I noticed Hasher doesn't call `__hash__` if the object being hashed implements it which I feel it should? If it did I could return a hash of the tokenizers.json file instead.\r\n\r\n```\r\n def __getstate__(self):\r\n \"\"\"\r\n Removes pre_tokenizer since it cannot be pickled\r\n \"\"\"\r\n logger.debug(\"Copy state dict\")\r\n self.backend_tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])\r\n out = copy.deepcopy(self.__dict__) #self.__dict__.copy()\r\n self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()\r\n\r\n return out\r\n```\r\n", "I'm glad you figured something out :)\r\n\r\nRegarding hashing: we're not using hashing for the same purpose as the python `__hash__` purpose (which is in general for dictionary lookups). For example it is allowed for python hashing to not return the same hash across sessions, while our hashing must return the same hashes across sessions for the caching to work properly." ]

[ "Hi ! For now this is probably the best option.\r\nWe might add a feature like this in the feature as well.\r\n\r\nDo you know any deduplication method that works on arbitrary big datasets without filling up RAM ?\r\nOtherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases", "Yes, I'd like to work on this feature once I'm done with #2500, but first I have to do some research, and see if the implementation wouldn't be too complex.\r\n\r\nIn the meantime, maybe [this lib](https://github.com/TomScheffers/pyarrow_ops) can help. However, note that this lib operates directly on pyarrow tables and relies only on `hash` to find duplicates (e.g. `-1` and `-2` have the same hash in Python 3, so this lib will treat them as duplicates), which doesn't make much sense.", "> Hi ! For now this is probably the best option.\r\n> We might add a feature like this in the feature as well.\r\n> \r\n> Do you know any deduplication method that works on arbitrary big datasets without filling up RAM ?\r\n> Otherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases\r\n\r\nGreat if this is can be done. Thanks!!\r\n\r\nNot sure if you are asking me. In any case I don't know of any unfortunately :( in practice if data is really large we normally do it with spark (only for info. I understand this is not useful in developing this library..)", "Hello,\r\n\r\nI'm also interested in this feature.\r\nHas there been progress on this issue?\r\n\r\nCould we use a similar trick as above, but with a better hashing algorithm like SHA?\r\n\r\nWe could also use a [bloom filter](https://en.wikipedia.org/wiki/Bloom_filter), should we care a lot about collision in this case?", "For reference, we can get a solution fairly easily if we assume that we can hold in memory all unique values. \r\n\r\n```python\r\nfrom datasets import Dataset\r\nfrom itertools import cycle\r\nfrom functools import partial\r\n\r\nmemory = set()\r\ndef is_unique(elem:Any , column: str, memory: set) -> bool:\r\n if elem[column] in memory:\r\n return False\r\n else:\r\n memory.add(elem[column])\r\n return True\r\n\r\n# Example dataset\r\nds = Dataset.from_dict({\"col1\" : [sent for i, sent in zip(range(10), cycle([\"apple\", \"orange\", \"pear\"]))],\r\n \"col2\": [i % 5 for i in range(10)]})\r\n\r\n# Drop duplicates in `ds` on \"col1\"\r\nds2 = ds.filter(partial(is_unique, column=\"col1\", memory=memory))\r\n```\r\n\r\nOf course, we can improve the API so that we can introduce `Dataset.drop_duplicates`.\r\nFor the parallel version, we can use a shared memory set.", "An approach that works assuming you can hold the all the unique document hashes in memory:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\ndef get_hash(example):\r\n \"\"\"Get hash of content field.\"\"\"\r\n return {\"hash\": hash(example[\"content\"])} # can use any hashing function here\r\n \r\ndef check_uniques(example, uniques):\r\n \"\"\"Check if current hash is still in set of unique hashes and remove if true.\"\"\"\r\n if example[\"hash\"] in uniques:\r\n uniques.remove(example[\"hash\"])\r\n return True\r\n else:\r\n return False\r\n\r\nds = load_dataset(\"some_dataset\")\r\nds = ds.map(get_hash)\r\nuniques = set(ds.unique(\"hash\"))\r\nds_filter = ds.filter(check_uniques, fn_kwargs={\"uniques\": uniques})\r\n```\r\nIf the `uniques` could be stored in arrow then no additional memory would used at all but I don't know if this is possible.\r\n", "@lvwerra hey, could you tell me how reliable is this deduplication method. i am currently using the same deduplication strategy to deduplicate a large text corpus to pretrain LLMs ~ 11B to 20B. just needed to ensure if this strategy would be fine on large datasets for LLMs pretraining. ", "Hi @StephennFernandes I'm also trying to pretrain an llm, and need to do deduplication for my dataset, \r\nwhich method you applied please?", "Hey @Manel-Hik \n\nThe following is a simpler yet really effective deduplication code that i has used in the past. \n\ngiven that I have limited training corpus for the languages I wanted to train i relied on this code. https://huggingface.co/datasets/Finnish-NLP/mc4_fi_cleaned/blob/main/deduplicate.py\n\n\nfor more robust and stronger deduplication, refer to this huggingface repo, that's newly released: https://github.com/huggingface/datatrove\n\n\n\n", "Thanks a lot Sure I will check it @StephennFernandes ", "Hi, is there any updates? Thanks!", "Update July 2024\r\n\r\nPyArrow now supports `first`/`last` aggregation which would allow us to implement this functionality. [Link](https://github.com/apache/arrow/issues/30950)\r\n\r\nso if we want to move in this direction we can :) Is that something we want to do? Would be happy to contribute." ]

[ "Hi @colbym-MM, thanks for reporting. We are fixing it." ]

[ "Sorry, I was using an old version of sequeval" ]

[ "Update on this: I'm computing the checksums of the data files. It will be available soon", "Added in #2575 :)" ]

[ "Hi ! Is this folder structure a standard, a bit like imagenet ?\r\nIn this case maybe we can consider having a dataset loader for cifar-like, imagenet-like, squad-like, conll-like etc. datasets ?\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nmy_custom_cifar = load_dataset(\"cifar_like\", data_dir=\"path/to/data/dir\")\r\n```\r\n\r\nLet me know what you think", "Yep that would be sweet - closing for now as we found a workaround. ", "@lhoestq I think we'll want a generic `image-folder` dataset (same as 'imagenet-like'). This is like `torchvision.datasets.ImageFolder`, and is something vision folks are used to seeing.", "Opening this back up, since I'm planning on tackling this. Already posted a quick version of it on my account on the hub.\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nds = load_dataset('nateraw/image-folder', data_files='PetImages/')\r\n```", "Bumping this one following our recent discussion @mariosasko @nateraw :)" ]

[ "Hi @arnaudstiegler, thanks for reporting. I'm investigating it.", "@arnaudstiegler I have just checked that these mismatches are already present in the original dataset: https://github.com/megagonlabs/SubjQA\r\n\r\nWe are going to contact the dataset owners to report this.", "I have:\r\n- opened an issue in their repo: https://github.com/megagonlabs/SubjQA/issues/3\r\n- written an email to all the paper authors", "Please [see my response](https://github.com/megagonlabs/SubjQA/issues/3#issuecomment-905160010). There will be a fix in a couple of days." ]

[ "👀 @TalSchuster", "Thanks @VictorSanh!\r\nThere's also a [notebook](https://aka.ms/python_puzzles) and [demo](https://aka.ms/python_puzzles_study) available now to try out some of the puzzles" ]

[ "That’s interesting thanks, let’s see what we can do. Can you detail your last sentence? I’m not sure I understand it well.", "Hi ! I just re-ran a quick benchmark and using `to_numpy()` seems to be faster now:\r\n\r\n```python\r\nimport pyarrow as pa # I used pyarrow 3.0.0\r\nimport numpy as np\r\n\r\nn, max_length = 1_000, 512\r\nlow, high, size = 0, 2 << 16, (n, max_length)\r\n\r\ntable = pa.Table.from_pydict({\r\n \"input_ids\": np.random.default_rng(42).integers(low=low, high=high, size=size).tolist()\r\n})\r\n\r\n\r\n%%timeit\r\n_ = table.to_pandas()[\"input_ids\"].to_numpy()\r\n# 1.44 ms ± 80.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)\r\n\r\n%%timeit\r\n_ = table[\"input_ids\"].to_pandas().to_numpy()\r\n# 461 µs ± 14.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)\r\n\r\n%%timeit\r\n_ = table[\"input_ids\"].to_numpy()\r\n# 317 µs ± 5.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)\r\n```\r\n\r\nCurrently the conversion from arrow to numpy is done in the NumpyArrowExtractor here:\r\n\r\nhttps://github.com/huggingface/datasets/blob/d6d0ede9486ffad7944642ca9a326e058b676788/src/datasets/formatting/formatting.py#L143-L166\r\n\r\nLet's update the NumpyArrowExtractor to call `to_numpy` directly and see how our github benchmarks evolve ?__", "Sounds like a plan @lhoestq If you create a PR I'll pick it up and try it out right away! ", "@lhoestq I can also prepare the PR, just lmk. ", "I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing?", "I created https://github.com/huggingface/datasets/pull/2505 if you want to play with it @vblagoje ", "> I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing?\r\n\r\n@thomwolf starting from the top, each rectangle represents the cumulative amount of it takes to execute the method call. Therefore, format_batch in torch_formatter.py takes ~20 sec, and the largest portion of that call is taken by to_pandas call and the smaller portion (grey rectangle) by the other method invocation(s) in format_batch (series_to_numpy etc). \r\n\r\nFeatures of the dataset are BERT pre-training model input columns i.e:\r\n```\r\nf = Features({ \r\n \"input_ids\": Sequence(feature=Value(dtype=\"int32\")), \r\n \"attention_mask\": Sequence(feature=Value(dtype=\"int8\")), \r\n \"token_type_ids\": Sequence(feature=Value(dtype=\"int8\")), \r\n \"labels\": Sequence(feature=Value(dtype=\"int32\")), \r\n \"next_sentence_label\": Value(dtype=\"int8\")\r\n})\r\n```\r\n\r\nI'll work with @lhoestq till we get to the bottom of this one. \r\n ", "@lhoestq the proposed branch is faster, but overall training speedup is a few percentage points. I couldn't figure out how to include the GitHub branch into setup.py, so I couldn't start NVidia optimized Docker-based pre-training run. But on bare metal, there is a slight improvement. I'll do some more performance traces. ", "Hi @vblagoje, to install Datasets from @lhoestq PR reference #2505, you can use:\r\n```shell\r\npip install git+ssh://[email protected]/huggingface/datasets.git@refs/pull/2505/head#egg=datasets\r\n```", "Hey @albertvillanova yes thank you, I am aware, I can easily pull it from a terminal command line but then I can't automate docker image builds as dependencies are picked up from setup.py and for some reason setup.py doesn't accept this string format.", "@vblagoje in that case, you can add this to your `setup.py`:\r\n```python\r\n install_requires=[\r\n \"datasets @ git+ssh://[email protected]/huggingface/datasets.git@refs/pull/2505/head\",\r\n```", "@lhoestq @thomwolf @albertvillanova The new approach is definitely faster, dataloader now takes less than 3% cumulative time (pink rectangle two rectangles to the right of tensor.py backward invocation)\r\n\r\n\r\n\r\nWhen we drill down into dataloader next invocation we get:\r\n\r\n\r\n\r\nAnd finally format_batch:\r\n\r\n\r\n\r\n\r\nNot sure this could be further improved but this is definitely a decent step forward.\r\n\r\n", "> ```python\r\n> datasets @ git+ssh://[email protected]/huggingface/datasets.git@refs/pull/2505/head\r\n> ```\r\n\r\n@albertvillanova how would I replace datasets dependency in https://github.com/huggingface/transformers/blob/master/setup.py as the above approach is not working. ", "@vblagoje I tested my proposed approach before posting it here and it worked for me. \r\n\r\nIs it not working in your case because of the SSH protocol? In that case you could try the same approach but using HTTPS:\r\n```\r\n\"datasets @ git+https://github.com/huggingface/datasets.git@refs/pull/2505/head\",\r\n``` ", "Also note the blanks before and after the `@`.", "@albertvillanova of course it works. Apologies. I needed to change datasets in all deps references , like [here](https://github.com/huggingface/transformers/blob/master/setup.py#L235) for example. ", "Is time spent casting an issue here? See https://github.com/huggingface/datasets/issues/4676 that Datasets can spend huge amounts of time repeatedly casting to Python objects." ]

[]

[]

[ "Hi @albertvillanova, I hope you are doing well.\r\n\r\nI am interested in this issue, is this still unresolved and open ?\r\n\r\nThe link you have provided in the above message directs to a webpage that does not exist.\r\n\r\nThanks and Regards" ]

[]

[]

[ "#self-assign" ]

[ "My suggestion for this would be to have this enabled by default.\r\n\r\nPlus I don't know if there should be a dedicated issue to that is another functionality. But I propose layered building rather than all at once. That is:\r\n\r\n1. uncompress a handful of files via a generator enough to generate one arrow file\r\n2. process arrow file 1\r\n3. delete all the files that went in and aren't needed anymore.\r\n\r\nrinse and repeat.\r\n\r\n1. This way much less disc space will be required - e.g. on JZ we won't be running into inode limitation, also it'd help with the collaborative hub training project\r\n2. The user doesn't need to go and manually clean up all the huge files that were left after pre-processing\r\n3. It would already include deleting temp files this issue is talking about\r\n\r\nI wonder if the new streaming API would be of help, except here the streaming would be into arrow files as the destination, rather than dataloaders." ]

[ "For example to be able to send uncompressed and temp build files to another volume/partition, so that the user gets the minimal disk usage on their primary setup - and ends up with just the downloaded compressed data + arrow files, but outsourcing the huge files and building to another partition. e.g. on JZ there is a special partition for fast data, but it's also volatile, so only temp files should go there.\r\n\r\nThink of it as `TMPDIR` so we need the equivalent for `datasets`." ]

[ "I've aligned the release script with Transformers in #6004, so I think this issue can be closed." ]

[ "This bug was fixed in #1995. Upgrading datasets to version 1.6 fixes the issue!", "Indeed was a fixed bug.\r\nWorks on version 1.8\r\nThanks " ]

[ "Hi ! `load_from_disk` doesn't move the data. If you specify a local path to your mounted drive, then the dataset is going to be loaded directly from the arrow file in this directory. The cache files that result from `map` operations are also stored in the same directory by default.\r\n\r\nHowever note than writing data to your google drive actually fills the VM's disk (see https://github.com/huggingface/datasets/issues/643)\r\n\r\nGiven that, I don't think that changing the cache directory changes anything.\r\n\r\nLet me know what you think", "Thanks for your answer! I am a little surprised since I just want to read the dataset.\r\n\r\nAfter debugging a bit, I noticed that the VM’s disk fills up when the tables (generator) are converted to a list:\r\n\r\nhttps://github.com/huggingface/datasets/blob/5ba149773d23369617563d752aca922081277ec2/src/datasets/table.py#L850\r\n\r\nIf I try to iterate through the table’s generator e.g.: \r\n\r\n`length = sum(1 for x in tables)`\r\n\r\nthe VM’s disk fills up as well.\r\n\r\nI’m running out of Ideas 😄 ", "Indeed reading the data shouldn't increase the VM's disk. Not sure what google colab does under the hood for that to happen", "Apparently, Colab uses a local cache of the data files read/written from Google Drive. See:\r\n- https://github.com/googlecolab/colabtools/issues/2087#issuecomment-860818457\r\n- https://github.com/googlecolab/colabtools/issues/1915#issuecomment-804234540\r\n- https://github.com/googlecolab/colabtools/issues/2147#issuecomment-885052636" ]

[ "I have received a reply from Zenodo support:\r\n> We are currently investigating and fixing this issue related to GitHub releases. As soon as we have solved it we will reach back to you.", "Other repo maintainers had the same problem with Zenodo. \r\n\r\nThere is an open issue on their GitHub repo: zenodo/zenodo#2181", "I have received the following request from Zenodo support:\r\n> Could you send us the link to the repository as well as the release tag?\r\n\r\nMy reply:\r\n> Sure, here it is:\r\n> - Link to the repository: https://github.com/huggingface/datasets\r\n> - Link to the repository at the release tag: https://github.com/huggingface/datasets/releases/tag/1.8.0\r\n> - Release tag: 1.8.0", "Zenodo issue has been fixed. The 1.8.0 release DOI can be found here: https://zenodo.org/record/4946100#.YMd6vKj7RPY" ]

[]

[ "Hi ! It looks like the issue comes from pyarrow. What version of pyarrow are you using ? How did you install it ?", "Thank you for the quick reply! I have `pyarrow==4.0.0`, and I am installing with `pip`. It's not one of my explicit dependencies, so I assume it came along with something else.", "Could you trying reinstalling pyarrow with pip ?\r\nI'm not sure why it would check in your multicurtural-sc directory for source files.", "Sure! I tried reinstalling to get latest. pip was mad because it looks like Datasets currently wants <4.0.0 (which is interesting, because apparently I ended up with 4.0.0 already?), but I gave it a shot anyway:\r\n\r\n```bash\r\n$ pip install --upgrade --force-reinstall pyarrow\r\nCollecting pyarrow\r\n Downloading pyarrow-4.0.1-cp39-cp39-manylinux2014_x86_64.whl (21.9 MB)\r\n |████████████████████████████████| 21.9 MB 23.8 MB/s\r\nCollecting numpy>=1.16.6\r\n Using cached numpy-1.20.3-cp39-cp39-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (15.4 MB)\r\nInstalling collected packages: numpy, pyarrow\r\n Attempting uninstall: numpy\r\n Found existing installation: numpy 1.20.3\r\n Uninstalling numpy-1.20.3:\r\n Successfully uninstalled numpy-1.20.3\r\n Attempting uninstall: pyarrow\r\n Found existing installation: pyarrow 3.0.0\r\n Uninstalling pyarrow-3.0.0:\r\n Successfully uninstalled pyarrow-3.0.0\r\nERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\r\ndatasets 1.8.0 requires pyarrow<4.0.0,>=1.0.0, but you have pyarrow 4.0.1 which is incompatible.\r\nSuccessfully installed numpy-1.20.3 pyarrow-4.0.1\r\n```\r\n\r\nTrying it, the same issue:\r\n\r\n\r\n\r\nI tried installing `\"pyarrow<4.0.0\"`, which gave me 3.0.0. Running, still, same issue.\r\n\r\nI agree it's weird that pyarrow is checking the source code directory for its files. (There is no `pyarrow/` directory there.) To me, that makes it seem like an issue with how pyarrow is called.\r\n\r\nOut of curiosity, I tried running this with fewer workers to see when the error arises:\r\n\r\n- 1: ✅\r\n- 2: ✅\r\n- 4: ✅\r\n- 8: ✅\r\n- 10: ✅\r\n- 11: ❌ 🤔\r\n- 12: ❌\r\n- 16: ❌\r\n- 32: ❌\r\n\r\nchecking my datasets:\r\n\r\n```python\r\n>>> datasets\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['text'],\r\n num_rows: 389290\r\n })\r\n validation.sc: Dataset({\r\n features: ['text'],\r\n num_rows: 10 # 🤔\r\n })\r\n validation.wvs: Dataset({\r\n features: ['text'],\r\n num_rows: 93928\r\n })\r\n})\r\n```\r\n\r\nNew hypothesis: crash if `num_proc` > length of a dataset? 😅\r\n\r\nIf so, this might be totally my fault, as the caller. Could be a docs fix, or maybe this library could do a check to limit `num_proc` for this case?", "Good catch ! Not sure why it could raise such a weird issue from pyarrow though\r\nWe should definitely reduce num_proc to the length of the dataset if needed and log a warning.", "This has been fixed in #2566, thanks @connor-mccarthy !\r\nWe'll make a new release soon that includes the fix ;)" ]

[ "Any update on this? @lhoestq ", "Unless there is high demande I don't think we will end up implementing this. This is a lot of work with very few advantages" ]

[ "@VictorSanh , I think @mariosasko is already working on it. " ]

[ "cc: @krandiash (pinged in reverted PR)." ]

[ "Realized that `tensorflow_datasets` is not provided by Huggingface and should therefore raise the issue there." ]

[ "Hi ! The `blue` metric doesn't exist, but the `bleu` metric does.\r\nYou can get the full list of metrics [here](https://github.com/huggingface/datasets/tree/master/metrics) or by running\r\n```python\r\nfrom datasets import list_metrics\r\n\r\nprint(list_metrics())\r\n```", "Ah, my mistake. Thanks for correcting" ]

[ "Hi ! I'm pretty sure that the answers are not made available for the test set on purpose because it is part of the DynaBench benchmark, for which you can submit your predictions on the website.\r\nIn any case we should mention this in the dataset card of this dataset.", "Makes sense, but not intuitive for someone searching through the datasets. Thanks for adding the note to clarify." ]

[ "```\r\nFile \"/home/sasha/.local/share/virtualenvs/lib-ogGKnCK_/lib/python3.7/site-packages/streamlit/script_runner.py\", line 332, in _run_script\r\n exec(code, module.__dict__)\r\nFile \"/home/sasha/nlp-viewer/run.py\", line 233, in <module>\r\n configs = get_confs(option)\r\nFile \"/home/sasha/.local/share/virtualenvs/lib-ogGKnCK_/lib/python3.7/site-packages/streamlit/caching.py\", line 604, in wrapped_func\r\n return get_or_create_cached_value()\r\nFile \"/home/sasha/.local/share/virtualenvs/lib-ogGKnCK_/lib/python3.7/site-packages/streamlit/caching.py\", line 588, in get_or_create_cached_value\r\n return_value = func(*args, **kwargs)\r\nFile \"/home/sasha/nlp-viewer/run.py\", line 148, in get_confs\r\n builder_cls = nlp.load.import_main_class(module_path[0], dataset=True)\r\nFile \"/home/sasha/.local/share/virtualenvs/lib-ogGKnCK_/lib/python3.7/site-packages/datasets/load.py\", line 85, in import_main_class\r\n module = importlib.import_module(module_path)\r\nFile \"/usr/lib/python3.7/importlib/__init__.py\", line 127, in import_module\r\n return _bootstrap._gcd_import(name[level:], package, level)\r\nFile \"<frozen importlib._bootstrap>\", line 1006, in _gcd_import\r\nFile \"<frozen importlib._bootstrap>\", line 983, in _find_and_load\r\nFile \"<frozen importlib._bootstrap>\", line 967, in _find_and_load_unlocked\r\nFile \"<frozen importlib._bootstrap>\", line 677, in _load_unlocked\r\nFile \"<frozen importlib._bootstrap_external>\", line 728, in exec_module\r\nFile \"<frozen importlib._bootstrap>\", line 219, in _call_with_frames_removed\r\nFile \"/home/sasha/.cache/huggingface/modules/datasets_modules/datasets/yelp_polarity/a770787b2526bdcbfc29ac2d9beb8e820fbc15a03afd3ebc4fb9d8529de57544/yelp_polarity.py\", line 36, in <module>\r\n from datasets.tasks import TextClassification\r\n```", "Solved by updating the `nlpviewer`" ]

[ "Hi,\r\n\r\nThis is a known issue. More info on this issue can be found in #2061. If you are looking for an open-source contribution, there are step-by-step instructions in the linked issue that you can follow to fix it.", "Closed by #2466." ]

[ "Hi ! That would be nice indeed to at least have a warning, since we don't handle the max path length limit.\r\nAlso if we could have an error instead of an infinite loop I'm sure windows users will appreciate that", "Unfortunately, I know this problem very well... 😅 \r\n\r\nI remember having proposed to throw an error instead of hanging in an infinite loop #2220: 60c7d1b6b71469599a27147a08100f594e7a3f84, 8c8ab60018b00463edf1eca500e434ff061546fc \r\nbut @lhoestq told me:\r\n> Note that the filelock module comes from this project that hasn't changed in years - while still being used by ten of thousands of projects:\r\nhttps://github.com/benediktschmitt/py-filelock\r\n> \r\n> Unless we have proper tests for this, I wouldn't recommend to change it\r\n\r\nI opened an Issue requesting a warning/error at startup for that case: #2224", "@albertvillanova Thanks for additional info on this issue.\r\n\r\nYes, I think the best option is to throw an error instead of suppressing it in a loop. I've considered 2 more options, but I don't really like them:\r\n1. create a temporary file with a filename longer than 255 characters on import; if this fails, long paths are not enabled and raise a warning. I'm not sure about this approach because I don't like the idea of creating a temporary file on import for this purpose.\r\n2. check if long paths are enabled with [this code](https://stackoverflow.com/a/46546731/14095927). As mentioned in the comment, this code relies on an undocumented function and Win10-specific." ]

[ "Hi ! In your dataset script you must be yielding examples like\r\n```python\r\nfor line in file:\r\n ...\r\n yield key, {...}\r\n```\r\n\r\nSince `datasets` 1.7.0 we enforce the keys to be unique.\r\nHowever it looks like your examples generator creates duplicate keys: at least two examples have key 0.\r\n\r\nYou can fix that by making sure that your keys are unique.\r\n\r\nFor example if you use a counter to define the key of each example, make sure that your counter is not reset to 0 in during examples generation (between two open files for examples).\r\n\r\nLet me know if you have other questions :)", "Yup, I indeed was generating duplicate keys. Fixed it and now it's working." ]

[ "The tagger also doesn't insert the value for the `size_categories` field automatically, so this should be fixed too", "Thanks for reporting. Indeed the `extended` tag doesn't exist. Not sure why we had that in the tagger.\r\nThe repo of the tagger is here if someone wants to give this a try: https://github.com/huggingface/datasets-tagging\r\nOtherwise I can probably fix it next week", "I've opened a PR on `datasets-tagging` to fix the issue 🚀 ", "thanks ! this is fixed now" ]

[ "this is also the case for the following datasets and configurations:\r\n\r\n* `mlqa` with config `mlqa-translate-train.ar`\r\n\r\n", "The current task API is somewhat deprecated (we plan to align it with `train eval index` at some point), so I think we can close this issue." ]

[ "Thanks for reporting !\r\n#2433 fixed the issue, thanks @mariosasko :)\r\n\r\nWe'll do a patch release soon of the library.\r\nIn the meantime, you can use the fixed version of adversarial_qa by adding `script_version=\"master\"` in `load_dataset`" ]

[ "It should probably work out of the box to save structured data. If you want to show an example we can help you.", "An example of a toy dataset is like:\r\n```json\r\n[\r\n {\r\n \"name\": \"mike\",\r\n \"friends\": [\r\n \"tom\",\r\n \"lily\"\r\n ],\r\n \"articles\": [\r\n {\r\n \"title\": \"aaaaa\",\r\n \"reader\": [\r\n \"tom\",\r\n \"lucy\"\r\n ]\r\n }\r\n ]\r\n },\r\n {\r\n \"name\": \"tom\",\r\n \"friends\": [\r\n \"mike\",\r\n \"bbb\"\r\n ],\r\n \"articles\": [\r\n {\r\n \"title\": \"xxxxx\",\r\n \"reader\": [\r\n \"tom\",\r\n \"qqqq\"\r\n ]\r\n }\r\n ]\r\n }\r\n]\r\n```\r\nWe can use the friendship relation to build a directional graph, and a user node can be represented using the articles written by himself. And the relationship between articles can be built when the article has read by the same user.\r\nThis dataset can be used to model the heterogeneous relationship between users and articles, and this graph can be used to build recommendation systems to recommend articles to the user, or potential friends to the user.", "Hi,\r\n\r\nyou can do the following to load this data into a `Dataset`:\r\n```python\r\nfrom datasets import Dataset\r\nexamples = [\r\n {\r\n \"name\": \"mike\",\r\n \"friends\": [\r\n \"tom\",\r\n \"lily\"\r\n ],\r\n \"articles\": [\r\n {\r\n \"title\": \"aaaaa\",\r\n \"reader\": [\r\n \"tom\",\r\n \"lucy\"\r\n ]\r\n }\r\n ]\r\n },\r\n {\r\n \"name\": \"tom\",\r\n \"friends\": [\r\n \"mike\",\r\n \"bbb\"\r\n ],\r\n \"articles\": [\r\n {\r\n \"title\": \"xxxxx\",\r\n \"reader\": [\r\n \"tom\",\r\n \"qqqq\"\r\n ]\r\n }\r\n ]\r\n }\r\n]\r\n\r\nkeys = examples[0].keys()\r\nvalues = [ex.values() for ex in examples]\r\ndataset = Dataset.from_dict({k: list(v) for k, v in zip(keys, zip(*values))})\r\n```\r\n\r\nLet us know if this works for you.", "Thank you so much, and that works! I also have a question that if the dataset is very large, that cannot be loaded into the memory. How to create the Dataset?", "If your dataset doesn't fit in memory, store it in a local file and load it from there. Check out [this chapter](https://huggingface.co/docs/datasets/master/loading_datasets.html#from-local-files) in the docs for more info.", "Nice! Thanks for your help." ]

[ "Hi,\r\n\r\n`load_dataset` returns an instance of `DatasetDict` if `split` is not specified, so instead of `Dataset.load_from_disk`, use `DatasetDict.load_from_disk` to load the dataset from disk.", "Thanks it worked!", "Though I see a stream of issues open by people lost between datasets and datasets dicts so maybe there is here something that could be better in terms of UX. Could be better error handling or something else smarter to even avoid said errors but maybe we should think about this. Reopening to use this issue as a discussion place but feel free to open a new open if you prefer @lhoestq @albertvillanova ", "We should probably improve the error message indeed.\r\n\r\nAlso note that there exists a function `load_from_disk` that can load a Dataset or a DatasetDict. Under the hood it calls either `Dataset.load_from_disk` or `DatasetDict.load_from_disk`:\r\n\r\n\r\n```python\r\nfrom datasets import load_from_disk\r\n\r\ndataset_dict = load_from_disk(\"path/to/dataset/dict\")\r\nsingle_dataset = load_from_disk(\"path/to/single/dataset\")\r\n```", "I just opened #2437 to improve the error message", "Superseded by #2462 " ]

[ "It actually seems to happen all the time in above configuration:\r\n* the function `filter_by_duration` correctly loads cached processed dataset\r\n* the function `prepare_dataset` is always reexecuted\r\n\r\nI end up solving the issue by saving to disk my dataset at the end but I'm still wondering if it's a bug or limitation here.", "Hi ! The hash used for caching `map` results is the fingerprint of the resulting dataset. It is computed using three things:\r\n- the old fingerprint of the dataset\r\n- the hash of the function\r\n- the hash of the other parameters passed to `map`\r\n\r\nYou can compute the hash of your function (or any python object) with\r\n```python\r\nfrom datasets.fingerprint import Hasher\r\n\r\nmy_func = lambda x: x + 1\r\nprint(Hasher.hash(my_func))\r\n```\r\n\r\nIf `prepare_dataset` is always executed, maybe this is because your `processor` has a different hash each time you want to execute it.", "> If `prepare_dataset` is always executed, maybe this is because your `processor` has a different hash each time you want to execute it.\r\n\r\nYes I think that was the issue.\r\n\r\nFor the hash of the function:\r\n* does it consider just the name or the actual code of the function\r\n* does it consider variables that are not passed explicitly as parameters to the functions (such as the processor here)", "> does it consider just the name or the actual code of the function\r\n\r\nIt looks at the name and the actual code and all variables such as recursively. It uses `dill` to do so, which is based on `pickle`.\r\nBasically the hash is computed using the pickle bytes of your function (computed using `dill` to support most python objects).\r\n\r\n> does it consider variables that are not passed explicitly as parameters to the functions (such as the processor here)\r\n\r\nYes it does thanks to recursive pickling.", "Thanks for these explanations. I'm closing the issue." ]

[ "Hi ! Can you try using a more up-to-date version ? We added the task_templates in `datasets` 1.7.0.\r\n\r\nIdeally when you're working on new datasets, you should install and use the local version of your fork of `datasets`. Here I think you tried to run the 1.7.0 tests with the 1.6.2 code" ]

[ "> I can provide a PR l8er...\r\n\r\nSee #2425 " ]

[ "Hi @cindyxinyiwang,\r\nDid you try adding `.arrow` after `cache_file_name` argument? Here I think they're expecting something like that only for a cache file:\r\nhttps://github.com/huggingface/datasets/blob/e08362256fb157c0b3038437fc0d7a0bbb50de5c/src/datasets/arrow_dataset.py#L1556-L1558", "Hi ! `cache_file_name` is an argument of the `Dataset.map` method. Can you check that your `dataset` is indeed a `Dataset` object ?\r\n\r\nIf you loaded several splits, then it would actually be a `DatasetDict` (one dataset per split, in a dictionary).\r\nIn this case, since there are several datasets in the dict, the `DatasetDict.map` method requires a `cache_file_names` argument (with an 's'), so that you can provide one file name per split.", "I think you are right. I used cache_file_names={data1: name1, data2: name2} and it works. Thank you!" ]

[]

[]

[ "I faced the similar problem. Downgrading datasets to 1.5.0 fixed it.", "Thanks for reporting, I'm looking into it", "I just opened #2438 to fix this :)", "Hi ! This has been fixed in the 1.8.0 release of `datasets`" ]

[ "Hi,\r\n\r\ndid you fill out the env info section manually or by copy-pasting the output of the `datasets-cli env` command?\r\n\r\nThis code should work without issues on 1.6.2 version (I'm working on master (1.6.2.dev0 version) and can't reproduce this error).", "@mariosasko you are right I was still on `1.5.0`. " ]

[ "These ranges seem to be valid English. Closing." ]

[ "Hi ! Thanks for reporting\r\ncc @pjox is this an issue from the data ?\r\n\r\nAnyway we should at least mention that OSCAR could contain such contents in the dataset card, you're totally right @jerryIsHere ", "Hi @jerryIsHere , sorry for the late response! Sadly this is normal, the problem comes form fasttext's classifier which we used to create the original corpus. In general the classifier is not really capable of properly recognizing Yue Chineese so the file ends un being just noise from Common Crawl. Some of these problems with OSCAR were already discussed [here](https://arxiv.org/pdf/2103.12028.pdf) but we are working on explicitly documenting the problems by language on our website. In fact, could please you open an issue on [our repo](https://github.com/oscar-corpus/oscar-website/issues) as well so that we can track it?" ]

[ "That could be useful indeed! Feel free to open a PR on the dataset card if you already have some code that runs, otherwise we'll take care of it soon :) ", "I can open a PR but there is 2 details to fix:\r\n- the name for the corresponding key (e.g. `original_answer`)\r\n- how to implement it: I’m not sure what happens when you map `lambda x: {'input': ...}` as it keeps the other keys (e.g. `output`) intact but here since we want to set a nested value (e.g. `x['output']['original_answer']`) I implemented it with a regular function (not lambda), see below\r\n\r\n```py\r\ndef add_original_answer(x, trivia_qa, triviaqa_map):\r\n i = triviaqa_map[x['id']]\r\n x['output']['original_answer'] = trivia_qa['validation'][i]['answer']['value']\r\n return x\r\n```" ]

[]

[ "Looks like there are multiple issues regarding this (#2386, #2322) and it's a WIP #2329. Currently these datasets are being loaded in-memory which is causing this issue. Quoting @mariosasko here for a quick fix:\r\n\r\n> set `keep_in_memory` to `False` when loading a dataset (`sst = load_dataset(\"sst\", keep_in_memory=False)`) to prevent it from loading in-memory. Currently, in-memory datasets fail to find cached files due to this check (always False for them)\r\n\r\n", "Hi ! Since `datasets` 1.6.0 we no longer keep small datasets (<250MB) on disk and load them in RAM instead by default. This makes data processing and iterating on data faster. However datasets in RAM currently have no way to reload previous results from the cache (since nothing is written on disk). We are working on making the caching work for datasets in RAM.\r\n\r\nUntil then, I'd recommend passing `keep_in_memory=False` to the calls to `load_dataset` like here:\r\n\r\nhttps://github.com/huggingface/transformers/blob/223943872e8c9c3fc11db3c6e93da07f5177423f/examples/pytorch/language-modeling/run_clm.py#L233\r\n\r\nThis way you say explicitly that you want your dataset to stay on the disk, and it will be able to recover previously computed results from the cache.", "gotcha! thanks Quentin", "OK, It doesn't look like we can use the proposed workaround - see https://github.com/huggingface/transformers/issues/11801\r\n\r\nCould you please add an env var for us to be able to turn off this unwanted in our situation behavior? It is really problematic for dev work, when one needs to restart the training very often and needs a quick startup time. Manual editing of standard scripts is not a practical option when one uses examples.\r\n\r\nThis could also be a problem for tests, which will be slower because of lack of cache, albeit usually we use tiny datasets there. I think we want caching for tests.\r\n\r\nThank you.", "Hi @stas00, \r\n\r\nYou are right: an env variable is needed to turn off this behavior. I am adding it.\r\n\r\nFor the moment there is a config parameter to turn off this behavior: `datasets.config.MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES = None`\r\n\r\nYou can find this info in the docs:\r\n- in the docstring of the parameter `keep_in_memory` of the function [`load_datasets`](https://huggingface.co/docs/datasets/package_reference/loading_methods.html#datasets.load_dataset):\r\n- in a Note in the docs about [Loading a Dataset](https://huggingface.co/docs/datasets/loading_datasets.html#from-the-huggingface-hub)\r\n\r\n> The default in 🤗Datasets is to memory-map the dataset on drive if its size is larger than datasets.config.MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES (default 250 MiB); otherwise, the dataset is copied in-memory. This behavior can be disabled by setting datasets.config.MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES = None, and in this case the dataset is not loaded in memory.", "Yes, but this still requires one to edit the standard example scripts, so if I'm doing that already I just as well can add `keep_in_memory=False`.\r\n\r\nMay be the low hanging fruit is to add `MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES` env var to match the config, and if the user sets it to 0, then it'll be the same as `keep_in_memory=False` or `datasets.config.MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES=0`?", "@stas00, however, for the moment, setting the value to `0` is equivalent to the opposite, i.e. `keep_in_memory=True`. This means the max size until which I load in memory is 0 bytes.\r\n\r\nTell me if this is logical/convenient, or I should change it.", "In my PR, to turn off current default bahavior, you should set env variable to one of: `{\"\", \"OFF\", \"NO\", \"FALSE\"}`.\r\n\r\nFor example:\r\n```\r\nMAX_IN_MEMORY_DATASET_SIZE_IN_BYTES=\r\n```", "IMHO, this behaviour is not very intuitive, as 0 is a normal quantity of bytes. So `MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES=0` to me reads as don't cache ever.\r\n\r\nAlso \"SIZE_IN_BYTES\" that can take one of `{\"\", \"OFF\", \"NO\", \"FALSE\"}` is also quite odd.\r\n\r\nI think supporting a very simple `MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES` that can accept any numerical value to match the name of the variable, requires minimal logic and is very straightforward. \r\n\r\nSo if you could adjust this logic - then `MAX_IN_MEMORY_DATASET_SIZE_IN_BYTES=0` is all that's needed to not do in-memory datasets.\r\n\r\nDoes it make sense?", "I understand your point @stas00, as I am not very convinced with current implementation.\r\n\r\nMy concern is: which numerical value should then pass a user who wants `keep_in_memory=True` by default, independently of dataset size? Currently it is `0` for this case.", "That's a good question, and again the normal bytes can be used for that:\r\n```\r\nMAX_IN_MEMORY_DATASET_SIZE_IN_BYTES=1e12 # (~2**40)\r\n```\r\nSince it's unlikely that anybody will have more than 1TB RAM.\r\n\r\nIt's also silly that it uses BYTES and not MBYTES - that level of refinement doesn't seem to be of a practical use in this context.\r\n\r\nNot sure when it was added and if there are back-compat issues here, but perhaps it could be renamed `MAX_IN_MEMORY_DATASET_SIZE` and support 1M, 1G, 1T, etc. \r\n\r\nBut scientific notation is quite intuitive too, as each 000 zeros is the next M, G, T multiplier. Minus the discrepancy of 1024 vs 1000, which adds up. And it is easy to write down `1e12`, as compared to `1099511627776` (2**40). (`1.1e12` is more exact).\r\n", "Great! Thanks, @stas00.\r\n\r\nI am implementing your suggestion to turn off default value when set to `0`.\r\n\r\nFor the other suggestion (allowing different metric prefixes), I will discuss with @lhoestq to agree on its implementation.", "Awesome! Thank you, @albertvillanova!!!\r\n\r\n" ]

[ "Thanks @bhavitvyamalik for referencing the workaround. Setting `keep_in_memory=False` is working." ]

[]

[]

[ "Hi ! This is because we are actually using the arrow streaming format. We plan to switch to the arrow IPC format.\r\nMore info at #1933 ", "Not sure if this was resolved, but I am getting a similar error when trying to load a dataset.arrow file directly: `ArrowInvalid: Not an Arrow file`", "Since we're using the streaming format, you need to use `open_stream`:\r\n\r\n```python\r\nimport pyarrow as pa\r\n\r\ndef in_memory_arrow_table_from_file(filename: str) -> pa.Table:\r\n in_memory_stream = pa.input_stream(filename)\r\n opened_stream = pa.ipc.open_stream(in_memory_stream)\r\n pa_table = opened_stream.read_all()\r\n return pa_table\r\n\r\ndef memory_mapped_arrow_table_from_file(filename: str) -> pa.Table:\r\n memory_mapped_stream = pa.memory_map(filename)\r\n opened_stream = pa.ipc.open_stream(memory_mapped_stream)\r\n pa_table = opened_stream.read_all()\r\n return pa_table\r\n```", "> 由于我们使用流格式，因此您需要使用`open_stream`：\r\n> \r\n> ```python\r\n> import pyarrow as pa\r\n> \r\n> def in_memory_arrow_table_from_file(filename: str) -> pa.Table:\r\n> in_memory_stream = pa.input_stream(filename)\r\n> opened_stream = pa.ipc.open_stream(in_memory_stream)\r\n> pa_table = opened_stream.read_all()\r\n> return pa_table\r\n> \r\n> def memory_mapped_arrow_table_from_file(filename: str) -> pa.Table:\r\n> memory_mapped_stream = pa.memory_map(filename)\r\n> opened_stream = pa.ipc.open_stream(memory_mapped_stream)\r\n> pa_table = opened_stream.read_all()\r\n> return pa_table\r\n> ```\r\nThank you very much for providing the code that can read arrow file to pa_table and finally to dict, but how to implement the reverse process, how to restore a dict to arrow file?\r\n" ]

[ "Version below works, checked again in the docs, and data_files should be a path.\r\n```\r\nds = datasets.load_dataset('my_script.py', \r\n data_files='/data/dir/corpus.txt', \r\n cache_dir='.')\r\n```" ]

[ "Closing this issue as the `task_templates` API has been deprecated." ]

[]

[]

[]

[ "We now have the `align_labels_with_mapping` method in the API for this purpose." ]

[]

[ "Just in case, this is a workaround that I use in my code and it seems to do the job.\r\n\r\n```python\r\nif use_gpu_index:\r\n data[\"train\"]._indexes[\"text_emb\"].faiss_index = faiss.index_gpu_to_cpu(data[\"train\"]._indexes[\"text_emb\"].faiss_index)\r\n```" ]

[ "Hi ! With @bhavitvyamalik we discussed about having something like\r\n```python\r\nfrom datasets import load_dataset_card\r\n\r\ndataset_card = load_dataset_card(\"squad\")\r\nprint(dataset_card.metadata.pretty_name)\r\n# Stanford Question Answering Dataset (SQuAD)\r\nprint(dataset_card.metadata.languages)\r\n# [\"en\"]\r\n\r\n```\r\nWhat do you think ?\r\n\r\nI don't know if you already have a way to load the model tags in `transformers` but we can agree on the API to have something consistent.\r\n\r\nAlso note that the pretty name would only be used to show users something prettier than a dataset id, but in the end the source of truth will stay the dataset id (here `squad`).", "That works for me!", "maybe use the hub-backed dataset_info method? (so there's only one parser of README.md metadata)?", "What dataset_info method are you talking about @julien-c ? In `huggingface_hub` I can only see `model_info`.", "hmm the equivalent method in `datasets` (which could go into `huggingface_hub` at some point)", "Indeed, this info can now be fetched with `huggingface_hub.dataset_info`, so I think we can close this issue." ]