Datasets:

---

miesnerjacob

/

github-issues

like 0

[]

[]

[]

[]

[]

[ "Add new \"dataset\"? ;)", "Lol, too used to Transformers ;-)" ]

[]

[]

[]

[]

[]

[ "Hello @juliahane, which config of GLUE causes you trouble?\r\nThe URLs are defined in the dataset script source code: https://github.com/huggingface/datasets/blob/master/datasets/glue/glue.py" ]

[]

[]

[ "Thank you for the feedback! I updated the code. " ]

[]

[]

[ "Hi @yangp725,\r\nThe SWDA has been added very recently and has not been released yet, thus it is not available in the `1.2.0` version of 🤗`datasets`.\r\nYou can still access it by installing the latest version of the library (master branch), by following instructions in [this issue](https://github.com/huggingface/datasets/issues/1641#issuecomment-751571471).\r\nLet me know if this helps !", "Thanks @SBrandeis ,\r\nProblem solved by downloading and installing the latest version datasets." ]

[]

[]

[]

[]

[]

[ "Hi @ZurMaD!\r\nThanks for your interest in 🤗 `datasets`. Support for image datasets is at an early stage, with CIFAR-10 added in #1617 \r\nMNIST is also on the way: #1730 \r\n\r\nIf you feel like adding another image dataset, I would advise starting by reading the [ADD_NEW_DATASET.md](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md) guide. New datasets are always very much appreciated 🚀\r\n" ]

[]

[]

[ "Hi @ameet-1997,\r\nI think what you are looking for is the `concatenate_datasets` function: https://huggingface.co/docs/datasets/processing.html?highlight=concatenate#concatenate-several-datasets\r\n\r\nFor your use case, I would use the [`map` method](https://huggingface.co/docs/datasets/processing.html?highlight=concatenate#processing-data-with-map) to transform the SQuAD sentences and the `concatenate` the original and mapped dataset.\r\n\r\nLet me know If this helps!", "That's a great idea! Thank you so much!\r\n\r\nWhen I try that solution, I get the following error when I try to concatenate `datasets` and `modified_dataset`. I have also attached the output I get when I print out those two variables. Am I missing something?\r\n\r\nCode:\r\n``` python\r\ncombined_dataset = concatenate_datasets([datasets, modified_dataset])\r\n```\r\n\r\nError:\r\n```\r\nAttributeError: 'DatasetDict' object has no attribute 'features'\r\n```\r\n\r\nOutput:\r\n```\r\n(Pdb) datasets\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['attention_mask', 'input_ids', 'special_tokens_mask'],\r\n num_rows: 493\r\n })\r\n})\r\n(Pdb) modified_dataset\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['attention_mask', 'input_ids', 'special_tokens_mask'],\r\n num_rows: 493\r\n })\r\n})\r\n```\r\n\r\nThe error is stemming from the fact that the attribute `datasets.features` does not exist. Would it not be possible to use `concatenate_datasets` in such a case? Is there an alternate solution?", "You should do `combined_dataset = concatenate_datasets([datasets['train'], modified_dataset['train']])`\r\n\r\nDidn't we talk about returning a Dataset instead of a DatasetDict with load_dataset and no split provided @lhoestq? Not sure it's the way to go but I'm wondering if it's not simpler for some use-cases.", "> Didn't we talk about returning a Dataset instead of a DatasetDict with load_dataset and no split provided @lhoestq? Not sure it's the way to go but I'm wondering if it's not simpler for some use-cases.\r\n\r\nMy opinion is that users should always know in advance what type of objects they're going to get. Otherwise the development workflow on their side is going to be pretty chaotic with sometimes unexpected behaviors.\r\nFor instance is `split=` is not specified it's currently always returning a DatasetDict. And if `split=\"train\"` is given for example it's always returning a Dataset.", "Thanks @thomwolf. Your solution worked!" ]

[]

[]

[ "Upgrading tensorflow to version 2.4.0 solved the issue.", "I still have the same error even with TF 2.4.0.", "And I have the same error with TF 2.4.1. I believe this issue should be reopened. Any ideas?!", "I'm seeing the same issue with TF 2.4.1 when running the following in https://colab.research.google.com/github/huggingface/datasets/blob/master/notebooks/Overview.ipynb:\r\n```\r\n!pip install git+https://github.com/google-research/bleurt.git\r\nreferences = [\"foo bar baz\", \"one two three\"]\r\nbleurt_metric = load_metric('bleurt')\r\npredictions = [\"foo bar\", \"four five six\"]\r\nbleurt_metric.compute(predictions=predictions, references=references)\r\n```", "@aleSuglia @oscartackstrom - Are you getting the error when running your code in a Jupyter notebook ?\r\n\r\nI tried reproducing this error again, and was unable to do so from the python command line console in a virtual environment similar to the one I originally used (and unfortunately no longer have access to) when I first got the error. \r\nHowever, I've managed to reproduce the error by running the same code in a Jupyter notebook running a kernel from the same virtual environment.\r\nThis made me suspect that the problem is somehow related to the Jupyter notebook.\r\n\r\nMore environment details:\r\n```\r\nOS: Ubuntu Linux 18.04\r\nconda==4.8.3\r\npython==3.8.5\r\ndatasets==1.3.0\r\ntensorflow==2.4.0\r\nBLEURT==0.0.1\r\nnotebook==6.2.0\r\n```", "This happens when running the notebook on colab. The issue seems to be that colab populates sys.argv with arguments not handled by bleurt.\r\n\r\nRunning this before calling bleurt fixes it:\r\n```\r\nimport sys\r\nsys.argv = sys.argv[:1]\r\n```\r\n\r\nNot the most elegant solution. Perhaps it needs to be fixed in the bleurt code itself rather than huggingface?\r\n\r\nThis is the output of `print(sys.argv)` when running on colab:\r\n```\r\n['/usr/local/lib/python3.7/dist-packages/ipykernel_launcher.py', '-f', '/root/.local/share/jupyter/runtime/kernel-a857a78c-44d6-4b9d-b18a-030b858ee327.json']\r\n```", "I got the error when running it from the command line. It looks more like an error that should be fixed in the BLEURT codebase.", "Seems to be a known issue in the bleurt codebase: https://github.com/google-research/bleurt/issues/24.", "Hi, the problem should be solved now." ]

[]

[]

[ "It's maybe a bit annoying to add but could we maybe have as well a version of the local data loading scripts in the package?\r\nThe `text`, `json`, `csv`. Thinking about people like in #1725 who are expecting to be able to work with local data without downloading anything.\r\n\r\nMaybe we can add them to package_data or something?", "Yes I mentioned this in #824 as well. I'm looking into it", "Alright now `csv`, `json`, `text` and `pandas` are \"packaged datasets\", i.e. they're part of the `datasets` package, which makes them available in offline mode without any change in terms of API:\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nd = load_dataset(\"csv\", data_files=[\"path/to/data.csv\"])\r\n```\r\n\r\nInstead of loading the dataset script from the module cache, it's loaded from inside the `datasets` package.\r\n\r\nI updated the test to still be able to fetch the dummy data files for those datasets from `datasets/{text|csv|pandas|json}/dummy` in the repo.", "Alright now all test pass :)\r\n(I don't thank you windows)", "LGTM! Since you're getting the local script's last modification date anyways do you think it might be a good idea to show it in the warning?", "> LGTM! Since you're getting the local script's last modification date anyways do you think it might be a good idea to show it in the warning?\r\n\r\nYep good idea. I added the date in the warning. For example `(last modified on Mon Nov 30 11:01:56 2020)`" ]

[]

[]

[ "You should rephrase your question or give more examples and details on what you want to do.\r\n\r\nit’s not possible to understand it and help you with only this information.", "sorry for that.\r\ni want to know how could i load the train set and the test set from the local ,which api or function should i use .\r\n", "Did you try to follow the instructions in the documentation?\r\nHere: https://huggingface.co/docs/datasets/loading_datasets.html#from-local-files", "thanks a lot \r\ni find that the problem is i dont use vpn...\r\nso i have to keep my net work even if i want to load the local data ?", "We will solve this soon (cf #1724)", "thanks a lot" ]

[]

[]

[ "I created the documentation for `FileSystem Integration for cloud storage` with loading and saving datasets to/from a filesystem with an example of using `datasets.filesystem.S3Filesystem`. I added a note on the `Saving a processed dataset on disk and reload` saying that it is also possible to use other filesystems and cloud storages such as S3 with a link to the newly created documentation page from me. \r\nI Attach a screenshot of it here. \r\n\r\n" ]

[]

[]

[ "Transferred to `datasets` based on the stack trace.", "Hi @lkcao !\r\nYour issue is indeed related to `datasets`. In addition to installing the package manually, you will need to download the `text.py` script on your server. You'll find it (under `datasets/datasets/text`: https://github.com/huggingface/datasets/blob/master/datasets/text/text.py.\r\nThen you can change the line 221 of `run_mlm_new.py` into:\r\n```python\r\n datasets = load_dataset('/path/to/text.py', data_files=data_files)\r\n```\r\nWhere `/path/to/text.py` is the path on the server where you saved the `text.py` script.", "We're working on including the local dataset builders (csv, text, json etc.) directly in the `datasets` package so that they can be used offline", "The local dataset builders (csv, text , json and pandas) are now part of the `datasets` package since #1726 :)\r\nYou can now use them offline\r\n```python\r\ndatasets = load_dataset('text', data_files=data_files)\r\n```\r\n\r\nWe'll do a new release soon", "> The local dataset builders (csv, text , json and pandas) are now part of the `datasets` package since #1726 :)\r\n> You can now use them offline\r\n> \r\n> ```python\r\n> datasets = load_dataset('text', data_files=data_files)\r\n> ```\r\n> \r\n> We'll do a new release soon\r\n\r\nso the new version release now?", "Yes it's been available since datasets 1.3.0 !" ]

[]

[]

[ "The current version of Wiki-Auto dataset contains a filtered version of the aligned dataset. The commit adds unfiltered versions of the data that can be useful the GEM task participants." ]

[]

[]

[ "> Nice !\r\n> \r\n> Could you try to reduce the size of the dummy_data.zip files ? they're quite big (300KB)\r\n\r\nYes, I think it might make sense to enhance the tool a tiny bit to prevent this automatically", "That's the lightest I can make it...it's long-range summarization so a single sample has ~11000 tokens. ", "Ok thanks :)", "Awesome good to merge for me :-) " ]

[]

[]

[ "Quick question, @lhoestq. In this specific dataset, two special types `GPE_LOC` and `GPE_ORG` can easily be altered depending on the task, choosing either the more general `GPE` tag or the more specific `LOC`/`ORG` tags, conflating them with the other annotations of the same type. However, I have not found an easy way to implement that. Using splits or configs does not seem appropriate.\r\n", "About the `GPE_LOC` and `GPE_ORG`. The original NorNE paper in which they published the dataset, does an evaluation on three different NER tag sets, one considering `GPE_LOC` and `GPE_ORG` as they are, another changing them to be just `GPE`, and another one by changing it to become `LOC` and `ORG`. The called these sets, `norne-full`, `norne-7`, and `norne-9`. What I would like is to provide a way for the user of this dataset to get `norne-7` and `norne-9` without having to duplicate the code.", "Ok I see !\r\nI guess you can have three configurations `norne-full`, `norne-7` and `norne-9`.\r\nEach config can have different feature types. You can simply check for the `self.config.name` in the `_info(self)` method and pick the right ClassLabel names accordingly. And then in `_generate_examples` as well you can check for `self.config.name` to know how to process the labels to yield either GPE_LOC/GPE_ORG, GPE or LOC/ORG", "But I'm already using the configurations for the different language\nvarieties. So you propose having something like `bokmaal`, `bokmaal-7`,\netc? Would there be a different way? If not, I'd be fine the corpus as it\nis until we come up with a solution. Thanks in any case.\n\n--\nSent using a cell-phone, so sorry for the typos and wrong auto-corrections.\n\nOn Tue, Jan 19, 2021, 4:56 PM Quentin Lhoest <[email protected]>\nwrote:\n\n> Ok I see !\n> I guess you can have three configurations norne-full, norne-7 and norne-9.\n> Each config can have different feature types. You can simply check for the\n> self.config.name in the _info(self) method and pick the right ClassLabel\n> names accordingly. And then in _generate_examples as well you can check\n> for self.config.name to know how to process the labels to yield either\n> GPE_LOC/GPE_ORG, GPE or LOC/ORG\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/datasets/pull/1720#issuecomment-762936612>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AABKLYOWNDBD76WZPJHFCWLS2WTTHANCNFSM4V6GSUQA>\n> .\n>\n", "The first option about having configurations like `bokmaal-7`, `bokmaal-9` etc. would definitely work.\r\n\r\nA second option would be to add a parameter `ner_tags_set` to `NorneConfig` and then one could load them with\r\n```python\r\nbokmaal_full = load_dataset(\"norne\", \"bokmaal\", ner_tags_set=\"norne-full\")\r\n```\r\nfor example.\r\n\r\nWhat do you think ?", "Hi @versae have you had a chance to consider one of the two options for the config ?\r\nI think both are ok but I have a small preference for the first one since it's simpler to implement.\r\n\r\nFeel free to ping me if you have questions or if I can help :) ", "Hi @lhoestq. Agree, option 1 seems easier to implement. Just haven't had bandwidth to get to it yet. Hopefully starting next week I'll be able to update the PR.", "Hi @versae ! Did you manage to add the configurations ? Let me know if we can help you on this", "Hi @lhoestq, I do actually have to code ready, just need to generate the dummy data for it. ", "One thing I don't know how to do is to make `_info(self)` return the different NER tags in its `DatasetInfo` object depending on the specific config.", "OK, I think it's ready now.", "Closing this one and opening a new one with a cleaner commit log.", "All set now in #2154." ]

[]

[]

[]

[]

[]

[ "Thanks for reporting !\r\nI was able to reproduce thanks to your code and find the origin of the bug.\r\nThe cache was not reusing the same file because one object was not deterministic. It comes from a conversion from `set` to `list` in the `datasets.arrrow_dataset.transmit_format` function, where the resulting list would not always be in the same order and therefore the function that computes the hash used by the cache would not always return the same result.\r\nI'm opening a PR to fix this.\r\n\r\nAlso we plan to do a new release in the coming days so you can expect the fix to be available soon.\r\nNote that you can still specify `cache_file_name=` in the second `map()` call to name the cache file yourself if you want to.", "Thanks for the fast reply, waiting for the fix :)\r\n\r\nI tried to use `cache_file_names` and wasn't sure how, I tried to give it the following:\r\n```\r\ntokenized_datasets = tokenized_datasets.map(\r\n group_texts,\r\n batched=True,\r\n num_proc=60,\r\n load_from_cache_file=True,\r\n cache_file_names={k: f'.cache/{str(k)}' for k in tokenized_datasets}\r\n)\r\n```\r\n\r\nand got an error:\r\n```\r\nmultiprocess.pool.RemoteTraceback:\r\n\"\"\"\r\nTraceback (most recent call last):\r\n File \"/venv/lib/python3.6/site-packages/multiprocess/pool.py\", line 119, in worker\r\n result = (True, func(*args, **kwds))\r\n File \"/venv/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 157, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/venv/lib/python3.6/site-packages/datasets/fingerprint.py\", line 163, in wrapper\r\n out = func(self, *args, **kwargs)\r\n File \"/venv/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1491, in _map_single\r\n tmp_file = tempfile.NamedTemporaryFile(\"wb\", dir=os.path.dirname(cache_file_name), delete=False)\r\n File \"/usr/lib/python3.6/tempfile.py\", line 690, in NamedTemporaryFile\r\n (fd, name) = _mkstemp_inner(dir, prefix, suffix, flags, output_type)\r\n File \"/usr/lib/python3.6/tempfile.py\", line 401, in _mkstemp_inner\r\n fd = _os.open(file, flags, 0o600)\r\nFileNotFoundError: [Errno 2] No such file or directory: '_00000_of_00060.cache/tmpsvszxtop'\r\n\"\"\"\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n File \"test.py\", line 48, in <module>\r\n cache_file_names={k: f'.cache/{str(k)}' for k in tokenized_datasets}\r\n File \"/venv/lib/python3.6/site-packages/datasets/dataset_dict.py\", line 303, in map\r\n for k, dataset in self.items()\r\n File \"/venv/lib/python3.6/site-packages/datasets/dataset_dict.py\", line 303, in <dictcomp>\r\n for k, dataset in self.items()\r\n File \"/venv/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1317, in map\r\n transformed_shards = [r.get() for r in results]\r\n File \"/venv/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1317, in <listcomp>\r\n transformed_shards = [r.get() for r in results]\r\n File \"/venv/lib/python3.6/site-packages/multiprocess/pool.py\", line 644, in get\r\n raise self._value\r\nFileNotFoundError: [Errno 2] No such file or directory: '_00000_of_00060.cache/tmpsvszxtop'\r\n```\r\n", "The documentation says\r\n```\r\ncache_file_names (`Optional[Dict[str, str]]`, defaults to `None`): Provide the name of a cache file to use to store the\r\n results of the computation instead of the automatically generated cache file name.\r\n You have to provide one :obj:`cache_file_name` per dataset in the dataset dictionary.\r\n```\r\nWhat is expected is simply the name of a file, not a path. The file will be located in the cache directory of the `wikitext` dataset. You can try again with something like\r\n```python\r\ncache_file_names = {k: f'tokenized_and_grouped_{str(k)}' for k in tokenized_datasets}\r\n```", "Managed to get `cache_file_names` working and caching works well with it\r\nHad to make a small modification for it to work:\r\n```\r\ncache_file_names = {k: f'tokenized_and_grouped_{str(k)}.arrow' for k in tokenized_datasets}\r\n```", "Another comment on `cache_file_names`, it doesn't save the produced cached files in the dataset's cache folder, it requires to give a path to an existing directory for it to work.\r\nI can confirm that this is how it works in `datasets==1.1.3`", "Oh yes indeed ! Maybe we need to update the docstring to mention that it is a path", "I fixed the docstring. Hopefully this is less confusing now: https://github.com/huggingface/datasets/commit/42ccc0012ba8864e6db1392430100f350236183a", "I upgraded to the latest version and I encountered some strange behaviour, the script I posted in the OP doesn't trigger recalculation, however, if I add the following change it does trigger partial recalculation, I am not sure if its something wrong on my machine or a bug:\r\n```\r\nfrom datasets import load_dataset\r\nfrom transformers import AutoTokenizer\r\n\r\ndatasets = load_dataset('wikitext', 'wikitext-103-raw-v1')\r\ntokenizer = AutoTokenizer.from_pretrained('bert-base-uncased', use_fast=True)\r\n\r\ncolumn_names = datasets[\"train\"].column_names\r\ntext_column_name = \"text\" if \"text\" in column_names else column_names[0]\r\ndef tokenize_function(examples):\r\n return tokenizer(examples[text_column_name], return_special_tokens_mask=True)\r\n# CHANGE\r\nprint('hello')\r\n# CHANGE\r\n\r\ntokenized_datasets = datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n...\r\n```\r\nI am using datasets in the `run_mlm.py` script in the transformers examples and I found that if I change the script without touching any of the preprocessing. it still triggers recalculation which is very weird\r\n\r\nEdit: accidently clicked the close issue button ", "This is because the `group_texts` line definition changes (it is defined 3 lines later than in the previous call). Currently if a function is moved elsewhere in a script we consider it to be different.\r\n\r\nNot sure this is actually a good idea to keep this behavior though. We had this as a security in the early development of the lib but now the recursive hashing of objects is robust so we can probably remove that.\r\nMoreover we're already ignoring the line definition for lambda functions.", "I opened a PR to change this, let me know what you think.", "Sounds great, thank you for your quick responses and help! Looking forward for the next release.", "I am having a similar issue where only the grouped files are loaded from cache while the tokenized ones aren't. I can confirm both datasets are being stored to file, but only the grouped version is loaded from cache. Not sure what might be going on. But I've tried to remove all kinds of non deterministic behaviour, but still no luck. Thanks for the help!\r\n\r\n\r\n```python\r\n # Datasets\r\n train = sorted(glob(args.data_dir + '*.{}'.format(args.ext)))\r\n if args.dev_split >= len(train):\r\n raise ValueError(\"Not enough dev files\")\r\n dev = []\r\n state = random.Random(1001)\r\n for _ in range(args.dev_split):\r\n dev.append(train.pop(state.randint(0, len(train) - 1)))\r\n\r\n max_seq_length = min(args.max_seq_length, tokenizer.model_max_length)\r\n\r\n def tokenize_function(examples):\r\n return tokenizer(examples['text'], return_special_tokens_mask=True)\r\n\r\n def group_texts(examples):\r\n # Concatenate all texts from our dataset and generate chunks of max_seq_length\r\n concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}\r\n total_length = len(concatenated_examples[list(examples.keys())[0]])\r\n # Truncate (not implementing padding)\r\n total_length = (total_length // max_seq_length) * max_seq_length\r\n # Split by chunks of max_seq_length\r\n result = {\r\n k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]\r\n for k, t in concatenated_examples.items()\r\n }\r\n return result\r\n\r\n datasets = load_dataset(\r\n 'text', name='DBNL', data_files={'train': train[:10], 'dev': dev[:5]}, \r\n cache_dir=args.data_cache_dir)\r\n datasets = datasets.map(tokenize_function, \r\n batched=True, remove_columns=['text'], \r\n cache_file_names={k: os.path.join(args.data_cache_dir, f'{k}-tokenized') for k in datasets},\r\n load_from_cache_file=not args.overwrite_cache)\r\n datasets = datasets.map(group_texts, \r\n batched=True,\r\n cache_file_names={k: os.path.join(args.data_cache_dir, f'{k}-grouped') for k in datasets},\r\n load_from_cache_file=not args.overwrite_cache)\r\n```\r\n\r\nAnd this is the log\r\n\r\n```\r\n04/26/2021 10:26:59 - WARNING - datasets.builder - Using custom data configuration DBNL-f8d988ad33ccf2c1\r\n04/26/2021 10:26:59 - WARNING - datasets.builder - Reusing dataset text (/home/manjavacasema/data/.cache/text/DBNL-f8d988ad33ccf2c1/0.0.0/e16f44aa1b321ece1f87b07977cc5d70be93d69b20486d6dacd62e12cf25c9a5)\r\n100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 13/13 [00:00<00:00, 21.07ba/s]\r\n100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 40/40 [00:01<00:00, 24.28ba/s]\r\n04/26/2021 10:27:01 - WARNING - datasets.arrow_dataset - Loading cached processed dataset at /home/manjavacasema/data/.cache/train-grouped\r\n04/26/2021 10:27:01 - WARNING - datasets.arrow_dataset - Loading cached processed dataset at /home/manjavacasema/data/.cache/dev-grouped\r\n```\r\n", "Hi ! What tokenizer are you using ?", "It's the ByteLevelBPETokenizer" ]

[]

[]

[ "Hi Dave,\r\nYou are more than welcome to open a PR to make these changes! 🤗\r\nYou will find the relevant information about opening a PR in the [contributing guide](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md) and in the [dataset addition guide](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).\r\n\r\nPinging also @lhoestq for the Google cloud matter.", "> I'd like to make a few minor changes, including the citation information and the `_URL` from which to download the dataset. Can I submit a PR for this?\r\n\r\nSure ! Also feel free to ping us for reviews or if we can help :)\r\n\r\n> It also looks like the dataset is being downloaded directly from Huggingface's Google cloud account rather than via the `_URL` in [scifact.py](https://github.com/huggingface/datasets/blob/master/datasets/scifact/scifact.py). Can you help me update the version on gcloud?\r\n\r\nWhat makes you think that ?\r\nAfaik there's no scifact on our google storage\r\n", "\r\n\r\n> > I'd like to make a few minor changes, including the citation information and the `_URL` from which to download the dataset. Can I submit a PR for this?\r\n> \r\n> Sure ! Also feel free to ping us for reviews or if we can help :)\r\n> \r\nOK! We're organizing a [shared task](https://sdproc.org/2021/sharedtasks.html#sciver) based on the dataset, and I made some updates and changed the download URL - so the current code points to a dead URL. I'll update appropriately once the task is finalized and make a PR.\r\n\r\n> > It also looks like the dataset is being downloaded directly from Huggingface's Google cloud account rather than via the `_URL` in [scifact.py](https://github.com/huggingface/datasets/blob/master/datasets/scifact/scifact.py). Can you help me update the version on gcloud?\r\n> \r\n> What makes you think that ?\r\n> Afaik there's no scifact on our google storage\r\n\r\nYou're right, I had the data cached on my machine somewhere. \r\n\r\n", "I opened a PR about this: https://github.com/huggingface/datasets/pull/1780. Closing this issue, will continue there." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Oh that's a really cool one, we'll review/merge it soon!\r\n\r\nIn the meantime, do you have any specific positive/negative feedback on the process of adding a datasets Max?\r\nDid you follow the instruction in the [detailed step-by-step](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md)?", "Thanks Thom, been a while, hope all is well!\r\n\r\nYes, I followed the step by step instructions and found them pretty straightforward. The only things I wasn't sure of were what should go into the YAML tags field for the dataset card, and whether there was a list of options somewhere (maybe akin to the metrics?) of the possible supported tasks. I found the rest very intuitive and the automated metadata and dummy data generation very handy. Thanks!", "Good point! pinging @yjernite here so he can improve this part!", "@maxbartolo cool addition!\r\n\r\nFor the YAML tag, you should use the tagging app we provide to choose from a drop-down menu:\r\nhttps://github.com/huggingface/datasets-tagging\r\n\r\nThe process is described toward the end of the [step-by-step guide](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md#tag-the-dataset-and-write-the-dataset-card), do you have any suggestions for making it easier to find?\r\n\r\nOtherwise, the dataset card is really cool, thanks for making it so complete!\r\n", "@yjernite\r\n\r\nThanks, YAML tags added. I think my main issue was with the flow of the [step-by-step guide](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md). For example, the [card creator](https://huggingface.co/datasets/card-creator/) is introduced in Step 4, right after creating an empty directory for your dataset. The first field it requires are the YAML tags, which (at least for me) was the last step of the process.\r\n\r\nI'd suggest having the guide structured in the same order as the creation process. For me it was something like:\r\n- Step 1: Preparing your env\r\n- Step 2: Write the loading/processing code\r\n- Step 3: Automatically generate dummy data and `dataset_infos.json`\r\n- Step 4: Tag the dataset\r\n- Step 5: Write the dataset card using the [card creator](https://huggingface.co/datasets/card-creator/)\r\n- Step 6: Open a Pull Request on the main HuggingFace repo and share your work!!\r\n\r\nThanks again!" ]

[]

[]

[ "Yes indeed the idea is to have the next release on conda cc @LysandreJik ", "Great! Did you guys have a timeframe in mind for the next release?\r\n\r\nThank you for all the great work in developing this library.", "I think we can have `datasets` on conda by next week. Will see what I can do!", "Thank you. Looking forward to it.", "`datasets` has been added to the huggingface channel thanks to @LysandreJik :)\r\nIt depends on conda-forge though\r\n\r\n```\r\nconda install -c huggingface -c conda-forge datasets\r\n```" ]

[]

[]

[ "When should we expect to see our dataset appear in the search dropdown at huggingface.co?", "Hi @eusip,\r\n\r\n> When should we expect to see our dataset appear in the search dropdown at huggingface.co?\r\n\r\nwhen this PR is merged.", "Thanks!", "I've implemented all the changes requested by @lhoestq but I made the mistake of trying to change the remote branch name. \r\n\r\nHopefully the changes are seen on your end as both branches `silicone` and `main` should be up-to-date.", "It looks like the PR includes changes about many other files than the ones for Silicone (+30,000 line changes)\r\n\r\nMaybe you can try to create another branch and another PR ?", "> It looks like the PR includes changes about many other files than the ones for Silicone (+30,000 line changes)\r\n> \r\n> Maybe you can try to create another branch and another PR ?\r\n\r\nSure. I will make a new pull request." ]

[]

[]

[]

[]

[]

[ "I haven't tested C4 on my side so there so there may be a few bugs in the code/adjustments to make.\r\nHere it looks like in c4.py, line 190 one of the `files_to_download` is `'/'` which is invalid.\r\nValid files are paths to local files or URLs to remote files." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Looks like we need to trim the ones with too many configs, will look into it tomorrow!" ]

[]

[]

[ "Hi ! Is this an issue you have with `openwebtext` specifically or also with other datasets ?\r\n\r\nIt looks like the downloaded file is corrupted and can't be extracted using `tarfile`.\r\nCould you try loading it again with \r\n```python\r\nimport datasets\r\ndatasets.load_dataset(\"openwebtext\", download_mode=\"force_redownload\")\r\n```" ]

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

[]

[]

[]

[]

[]

[]

[]

[ "I few comments here for discussion:\r\n- I'm not convinced yet the end user should really have to understand the difference between \"caching\" and 'fingerprinting\", what do you think? I think fingerprinting should probably stay as an internal thing. Is there a case where we want cahing without fingerprinting or vice-versa?\r\n- while I think the random fingerprint mechanism is smart, I have one question: when we disable caching or fingerprinting we also probably don't want the disk usage to grow so we should then try to keep only one cache file. Is it the case currently?\r\n- the warning should be emitted only once per session if possible (we have a mechanism to do that in transformers, you should ask Lysandre/Sylvain)\r\n\r\n", "About your points:\r\n- Yes I agree, I just wanted to bring the discussion on this point. Until now fingerprinting hasn't been blocking for user experience. I'll probably remove the enable/disable fingerprinting function to keep things simple from the user's perspective.\r\n- Right now every time a not in-place transform (i.e. map, filter) is applied, a new cache file is created. It is the case even if caching is disabled since disabling it only means that the cache file won't be reloaded. Therefore you're right that it might end up filling the disk with files that won't be reused. I like the idea of keeping only one cache file. Currently all the cache files are kept on disk until the user clears the cache. To be able to keep only one, we need to know if a dataset that has been transformed is still loaded or not. For example\r\n```python\r\n# case 1 - keep both cache files (dataset1 and dataset2)\r\ndataset2 = dataset1.map(...)\r\n# case 2 - keep only the new cache file\r\ndataset1 = dataset1.map(...)\r\n```\r\nIn python it doesn't seem trivial to detect such changes. One thing that we can actually do on the other hand is store the cache files in a temporary directory that is cleared when the session closes. I think that's a good a simple solution for this problem.\r\n- Yes good idea ! I don't like spam either :) ", "> * To be able to keep only one, we need to know if a dataset that has been transformed is still loaded or not. For example\r\n> \r\n> ```python\r\n> # case 1 - keep both cache files (dataset1 and dataset2)\r\n> dataset2 = dataset1.map(...)\r\n> # case 2 - keep only the new cache file\r\n> dataset1 = dataset1.map(...)\r\n> ```\r\n\r\nI see what you mean. It's a tricky question. One option would be that if caching is deactivated we have a single memory mapped file and have copy act as a copy by reference instead of a copy by value. We will then probably want a `copy()` or `deepcopy()` functionality. Maybe we should think a little bit about it though.", "- I like the idea of using a temporary directory per session!\r\n- If the default behavior when caching is disabled is to re-use the same file, I'm a little worried about people making mistakes and having to re-download and process from scratch.\r\n- So we already have a keyword argument for `dataset1 = dataset1.map(..., in_place=True)`?", "> * If the default behavior when caching is disabled is to re-use the same file, I'm a little worried about people making mistakes and having to re-download and process from scratch.\r\n\r\nWe should distinguish between the caching from load_dataset (base dataset cache files) and the caching after dataset transforms such as map or filter (transformed dataset cache files). When disabling caching only the second type (for map and filter) doesn't reload from cache files.\r\nTherefore nothing is re-downloaded. To re-download the dataset entirely the argument `download_mode=\"force_redownload\"` must be used in `load_dataset`.\r\nDo we have to think more about the naming to make things less confusing in your opinion ?\r\n\r\n> * So we already have a keyword argument for `dataset1 = dataset1.map(..., in_place=True)`?\r\n\r\nThere's no such `in_place` parameter in map, what do you mean exactly ?", "I updated the PR:\r\n- I removed the enable/disable fingerprinting function\r\n- if caching is disabled arrow files are written in a temporary directory that is deleted when session closes\r\n- the warning that is showed when hashing a transform fails is only showed once\r\n- I added the `set_caching_enabled` function to the docs and explained the caching mechanism and its relation with fingerprinting\r\n\r\nI would love to have some feedback :) ", "> > * So we already have a keyword argument for `dataset1 = dataset1.map(..., in_place=True)`?\r\n> \r\n> There's no such `in_place` parameter in map, what do you mean exactly ?\r\n\r\nSorry, that wasn't clear at all. I was responding to your previous comment about case 1 / case 2. I don't think the behavior should depend on the command, but we could have:\r\n\r\n```\r\n# case 1 - keep both cache files (dataset1 and dataset2)\r\ndataset2 = dataset1.map(...)\r\n# case 2 - keep only the new cache file\r\ndataset1 = dataset1.map(..., in_place=True)\r\n```\r\n\r\nCase 1 returns a new reference using the new cache file, case 2 returns the same reference", "> Sorry, that wasn't clear at all. I was responding to your previous comment about case 1 / case 2. I don't think the behavior should depend on the command, but we could have:\r\n> \r\n> ```\r\n> # case 1 - keep both cache files (dataset1 and dataset2)\r\n> dataset2 = dataset1.map(...)\r\n> # case 2 - keep only the new cache file\r\n> dataset1 = dataset1.map(..., in_place=True)\r\n> ```\r\n> \r\n> Case 1 returns a new reference using the new cache file, case 2 returns the same reference\r\n\r\nOk I see !\r\n`in_place` is a parameter that is used in general to designate a transform so I would name that differently (maybe `overwrite` or something like that).\r\nNot sure if it's possible to update an already existing arrow file that is memory-mapped, let me check real quick.\r\nAlso it's possible to call `dataset2.cleanup_cache_files()` to delete the other cache files if we create a new one after the transform. Or even to get the cache file with `dataset1.cache_files` and let the user remove them by hand.\r\n\r\nEDIT: updating an arrow file in place is not part of the current API of pyarrow, so we would have to make new files.\r\n" ]

[]

[]

[]

[]

[]

[ "Thanks for reporting.\r\nWe should indeed add all the missing dummy_data.zip and also the dataset_infos.json at least for lm1b, reclor and wikihow.\r\n\r\nFor c4 I haven't tested the script and I think we'll require some optimizations regarding beam datasets before processing it.\r\n" ]

[]

[]

[]

[]

[]

[ "not sure why the CI was not triggered though" ]

[]

[]

[ "Really cool!\r\n\r\nCan you add some task tags for `dialogue-modeling` (under `sequence-modeling`) and `parsing` (under `structured-prediction`)?" ]

[]

[]

[ "Same as #1698, can you add a task tag for dialogue-modeling (under sequence-modeling) :) ?" ]

[]

[]

[ "Maybe try to create a virtual env with python 3.8 or 3.7", "Thanks, @thomwolf! I fixed the issue by downgrading python to 3.7. ", "Damn sorry", "Damn sorry" ]

[]

[]

[ "> Thanks :)\r\n> \r\n> Apparently the dummy_data.zip got removed. Is this expected ?\r\n> Also can you remove the `data-pos.conll` file that you added ?\r\n\r\nNot expected. I forgot to remove the `dummy_data` folder used to create `dummy_data.zip`. \r\nChanged to only `dummy_data.zip`." ]

[]

[]

[ "Hi @lhoestq, on the OSCAR dataset, the document boundaries are defined by an empty line. Are there any chances to keep this empty line or explicitly group the sentences of a document? I'm asking for this 'cause I need to know if some sentences belong to the same document on my current OSCAR dataset usage.", "Indeed currently it yields one example per line and ignore the empty lines.\r\nMaybe the best is to group them by paragraph then, and yield one example when an empty line is found.\r\nWhat do you think ?", "I think to group them is the best choice indeed, I actually did this on [brwac](https://github.com/huggingface/datasets/tree/master/datasets/brwac) dataset too, it's another huge textual dataset.", "Ok I just launched the computation of the dataset_infos.json again by grouping lines in paragraphs.\r\nThe new _generate_examples is\r\n```python\r\n def _generate_examples(self, filepaths):\r\n \"\"\"This function returns the examples in the raw (text) form.\"\"\"\r\n id_ = 0\r\n current_lines = []\r\n for filepath in filepaths:\r\n logging.info(\"generating examples from = %s\", filepath)\r\n with gzip.open(filepath, \"rt\") as f:\r\n for line in f:\r\n if len(line.strip()) > 0:\r\n current_lines.append(line)\r\n else:\r\n feature = id_, {\"id\": id_, \"text\": \"\".join(current_lines)}\r\n yield feature\r\n id_ += 1\r\n current_lines = []\r\n # last paragraph\r\n if current_lines:\r\n feature = id_, {\"id\": id_, \"text\": \"\".join(current_lines)}\r\n yield feature\r\n```", "Is there any chance to also keep the sentences raw (without the `\"\".join()`)?. This is useful if you wanna train models where one of the tasks you perform is document sentence permutation... that's my case :)", "They are raw in the sense that nothing is changed from the raw file for each paragraph.\r\nYou can split sentences on new lines `\\n` for example.\r\n\r\nThe first example for the unshuffled deduplicated english is going to be \r\n> Mtendere Village was inspired by the vision of Chief Napoleon Dzombe, which he shared with John Blanchard during his first visit to Malawi. Chief Napoleon conveyed the desperate need for a program to intervene and care for the orphans and vulnerable children (OVC) in Malawi, and John committed to help.\r\n> Established in honor of John & Lindy’s son, Christopher Blanchard, this particular program is very dear to the Blanchard family. Dana Blanchard, or Mama Dana as she is more commonly referred to at Mtendere, lived on site during the initial development, and she returns each summer to spend the season with her Malawian family. The heart of the program is to be His hands and feet by caring for the children at Mtendere, and meeting their spiritual, physical, academic, and emotional needs.\r\n> [...]\r\n> 100X Development Foundation, Inc. is registered 503 (c)(3) nonprofit organization. Donations are deductable to the full extent allowable under IRS regulations.", "I thought the line reader would omit the `\\n` character. I can easily split the sentences as you suggested. Thanks @lhoestq! 😃 ", "The recomputation of the metadata finished a few days ago, I'll update the PR soon :) ", "Let me know if you have comments @pjox @jonatasgrosman :) \r\n\r\nOtherwise we can merge it", "Everything seems fine to me 😄 " ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "We could do something. There is a part of `.info` which is split specific (cache files, split instructions) but maybe if could be made to work.", "Yes this was kinda the idea I was going for. DatasetDict.info would be the shared info amongs the datasets (maybe even some info on how they differ). " ]

[]

[]

[ "Thanks for reporting, I opened a PR to fix that", "One the PR is merged the fix will be available in the next release of `datasets`.\r\n\r\nIf you don't want to wait the next release you can still load the script from the master branch with\r\n\r\n```python\r\nload_dataset(\"dane\", script_version=\"master\")\r\n```", "If you have other questions feel free to reopen :) " ]

[]

[]

[ "Thanks for reviewing and merging!" ]

[]

[]

[]

[]

[]

[ "Looks like the mapping function returns a dictionary with a 768-dim array in the `embeddings` field. Since the map is batched, we actually expect the `embeddings` field to be an array of shape (batch_size, 768) to have one embedding per example in the batch.\r\n\r\nTo fix that can you try to remove one of the `[0]` ? In my opinion you only need one of them, not two.", "It makes sense :D\r\n\r\nIt seems to work! Thanks a lot :))\r\n\r\nClosing the issue" ]

[]

[]

[]

[]

[]

[ "Hi @KennethEnevoldsen ,\r\nI think the issue might be that this dataset was added during the community sprint and has not been released yet. It will be available with the v2 of datasets.\r\nFor now, you should be able to load the datasets after installing the latest (master) version of datasets using pip:\r\npip install git+https://github.com/huggingface/datasets.git@master", "The `dane` dataset was added recently, that's why it wasn't available yet. We did an intermediate release today just before the v2.0.\r\n\r\nTo load it you can just update `datasets`\r\n```\r\npip install --upgrade datasets\r\n```\r\n\r\nand then you can load `dane` with\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\ndataset = load_dataset(\"dane\")\r\n```", "Thanks. Solved the problem." ]

[]

[]

[ "@lhoestq, can you please review this PR?", "Thanks for the suggestions. Updates were made and dataset_infos.json file was created again." ]

[]

[]

[ "cc100 was added recently, that's why it wasn't available yet.\r\n\r\nTo load it you can just update `datasets`\r\n```\r\npip install --upgrade datasets\r\n```\r\n\r\nand then you can load `cc100` with\r\n\r\n```python\r\nfrom datasets import load_dataset\r\n\r\nlang = \"en\"\r\ndataset = load_dataset(\"cc100\", lang=lang, split=\"train\")\r\n```" ]

[]

[]

[ "@lhoestq Thank you for your detailed comments! I fixed everything you suggested.\r\n\r\nPlease let me know if I'm missing anything else.", "It looks like the Transcript and Utterance objects are missing, maybe we can mention it in the README ? Or just add them ? @gmihaila @bhavitvyamalik ", "Hi @lhoestq,\r\nI'm working on this to add the full dataset", "> It looks like the Transcript and Utterance objects are missing, maybe we can mention it in the README ? Or just add them ? @gmihaila @bhavitvyamalik\r\n\r\n@lhoestq Any info on how to add them?", "@gmihaila, instead of using the current repo you should look into [this](https://github.com/cgpotts/swda). You can use the `csv` files uploaded in this repo (`swda.zip`) to access other fields and include them in this dataset. It has one dependency too, `swda.py`, you can download that separately and include it in your dataset's folder to be imported while reading the `csv` files.\r\n\r\nAlmost all the attributes of `Transcript` and `Utterance` objects are of the type str, int, or list. As far as `trees` attribute is concerned in utterance object you can simply parse it as string and user can maybe later convert it to nltk.tree object", "@bhavitvyamalik Thank you for the clarification! \r\n\r\nI didn't use [that](https://github.com/cgpotts/swda) because it doesn't have the splits. I think in combination with [what I used](https://github.com/NathanDuran/Switchboard-Corpus) would help.\r\n\r\nLet me know if I can help! I can make those changes if you don't have the time.", "I'm a bit busy for the next 2 weeks. I'll be able to complete it by end of January only. Maybe you can start with it and I'll help you?\r\nAlso, I looked into the official train/val/test splits and not all the files are there in the repo I used so I think either we'll have to skip them or put all of that into just train", "Yes, I can start working on it and ask you to do a code review.\r\n\r\nYes, not all files are there. I'll try to find papers that have the correct and full splits, if not, I'll do like you suggested.\r\n\r\nThank you again for your help @bhavitvyamalik !" ]

[]

[]