Datasets:

---

dongxiaoxia194

/

github-issues

like 0

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6898). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "After this PR, the README file looks like:\r\n\r\n\r\n" ]

[]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Hello, @bghira.\r\n\r\nThanks for reporting. Please note that the text originating the error is not supposed to be valid YAML: it contains the instructions to generate the actual YAML content, that should replace the instructions comment.\r\n\r\nOn the other hand, I agree that it is not nice to have that YAML error message at the top of the page: \r\n\r\n\r\nI am proposing a change to make the YAML error disappear.", "thanks albert! i looked at it for a while to figure it out. i think the `raw` view option is the correct way to look at it?" ]

[]

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6895). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

[ { "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" } ]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6883). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

[]

[]

[]

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

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[]

[]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "The error is caused by malformed basenames of the files within the TARs:\r\n- `15_Cohen_1-s2.0-S0929664620300449-gr3_lrg-b.png` becomes `15_Cohen_1-s2` as the grouping `__key__`, and `0-S0929664620300449-gr3_lrg-b.png` as the additional key to be added to the example\r\n- whereas the intended behavior was to use `15_Cohen_1-s2.0-S0929664620300449-gr3_lrg-b` as the grouping `__key__`, and `png` as the additional key to be added to the example\r\n\r\nTo get the expected behavior, the basenames of the files within the TARs should be fixed so that they only contain a single dot, the one separating the file extension.", "I reopen it because I think we should try to give a clearer error message with a specific error code.\r\n\r\nFor now, it's hard for the user to understand where the error comes from (not everybody knows the subtleties of the webdataset filename structure).\r\n\r\n(we can transfer it to https://github.com/huggingface/dataset-viewer if it fits better there)", "same with .jpg -> https://huggingface.co/datasets/ProGamerGov/synthetic-dataset-1m-dalle3-high-quality-captions\r\n\r\n```\r\nError code: DatasetGenerationError\r\nException: DatasetGenerationError\r\nMessage: An error occurred while generating the dataset\r\nTraceback: Traceback (most recent call last):\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1748, in _prepare_split_single\r\n for key, record in generator:\r\n File \"/src/services/worker/src/worker/job_runners/config/parquet_and_info.py\", line 818, in wrapped\r\n for item in generator(*args, **kwargs):\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/webdataset/webdataset.py\", line 109, in _generate_examples\r\n example[field_name] = {\"path\": example[\"__key__\"] + \".\" + field_name, \"bytes\": example[field_name]}\r\n KeyError: 'jpg'\r\n \r\n The above exception was the direct cause of the following exception:\r\n \r\n Traceback (most recent call last):\r\n File \"/src/services/worker/src/worker/job_runners/config/parquet_and_info.py\", line 1316, in compute_config_parquet_and_info_response\r\n parquet_operations, partial = stream_convert_to_parquet(\r\n File \"/src/services/worker/src/worker/job_runners/config/parquet_and_info.py\", line 909, in stream_convert_to_parquet\r\n builder._prepare_split(\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1627, in _prepare_split\r\n for job_id, done, content in self._prepare_split_single(\r\n File \"/src/services/worker/.venv/lib/python3.9/site-packages/datasets/builder.py\", line 1784, in _prepare_split_single\r\n raise DatasetGenerationError(\"An error occurred while generating the dataset\") from e\r\n datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset\r\n```\r\n", "More details in the spec (https://docs.google.com/document/d/18OdLjruFNX74ILmgrdiCI9J1fQZuhzzRBCHV9URWto0/edit#heading=h.hkptaq2kct2s)\r\n\r\n> The prefix of a file is all directory components of the file plus the file name component up to the first “.” in the file name.\r\n> The last extension (i.e., the portion after the last “.”) in a file name determines the file type.\r\n\r\n> Example:\r\n\timages17/image194.left.jpg\r\n\timages17/image194.right.jpg\r\n\timages17/image194.json\r\n\timages17/image12.left.jpg\r\n\timages17/image12.json\r\n\timages17/image12.right.jpg\r\n\timages3/image1459.left.jpg\r\n> \t…\r\n> When reading this with a WebDataset library, you would get the following two dictionaries back in sequence:\r\n\r\n { “__key__”: “images17/image194”, “left.jpg”: b”...”, “right.jpg”: b”...”, “json”: b”...”}\r\n { “__key__”: “images17/image12”, “left.jpg”: b”...”, “right.jpg”: b”...”, “json”: b”...”}\r\n", "OK, the issue is different in the latter case: some files are suffixed as `.jpeg`, and others as `.jpg` :)\r\n\r\nIs it a limitation of the webdataset format, or of the datasets library @lhoestq? And could we be able to give a clearer error?" ]

[]

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6878). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6876). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "transformers 4.40.2 was release yesterday but not sure if it contains the fix", "@lhoestq yes I knew transformers 4.40.2 was released yesterday, but I had checked that it does not contain the fix: only 2 bug fixes. That is why our CI continues failing in this PR. We will have to wait until the next minor version.", "> If we urgently need some dev feature for dataset-viewer, I would suggest pushing the feature (cherry-picked) to a dedicated branch with 2.19.1 as its starting point (without opening a PR), and install datasets from that branch.\r\n\r\nI have done so:\r\n- Created a branch from 2.19.1: https://github.com/huggingface/datasets/tree/datasets-2.19.1-hotfix\r\n- Cherry-picked the commit in this PR: https://github.com/huggingface/datasets/commit/3638183e2f7e0dce8924e46e7cc21bf6d5d7adfb\r\n- Opened a PR in dataset-viewer to update datasets to this revision: https://github.com/huggingface/dataset-viewer/pull/2783" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6874). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Before pandas-2.2.0, the function `ujson_loads` was named `loads`: https://github.com/pandas-dev/pandas/blob/v2.1.0/pandas/io/json/__init__.py#L5\r\n```python\r\nimport ujson_loads as loads\r\n```" ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[ "Thanks! Feel free to ping me for examples. May not respond immediately because we're all busy but would like to help.", "Hi @natolambert, could you please give some examples of JSON files to benchmark?\r\n\r\nPlease note that this JSON file (https://huggingface.co/datasets/allenai/reward-bench-results/blob/main/eval-set-scores/Ray2333/reward-model-Mistral-7B-instruct-Unified-Feedback.json) is not in \"records\" orient; instead it has the following structure:\r\n```json\r\n{\r\n \"chat_template\": \"tulu\",\r\n \"id\": [30, 34, 35,...],\r\n \"model\": \"Ray2333/reward-model-Mistral-7B-instruct-Unified-Feedback\",\r\n \"model_type\": \"Seq. Classifier\",\r\n \"results\": [1, 1, 1, ...],\r\n \"scores_chosen\": [4.421875, 1.8916015625, 3.8515625,...],\r\n \"scores_rejected\": [-2.416015625, -1.47265625, -0.9912109375,...],\r\n \"subset\": [\"alpacaeval-easy\", \"alpacaeval-easy\", \"alpacaeval-easy\",...]\r\n \"text_chosen\": [\"<s>[INST] How do I detail a...\",...],\r\n \"text_rejected\": [\"<s>[INST] How do I detail a...\",...]\r\n}\r\n```\r\n\r\nNote that \"records\" orient should be a list (not a dict) with each row as one item of the list:\r\n```json\r\n[\r\n {\"chat_template\": \"tulu\", \"id\": 30,... },\r\n {\"chat_template\": \"tulu\", \"id\": 34,... },\r\n ...\r\n]\r\n```", "We use a mix (which is a mess), here's an example with the records orient\r\nhttps://huggingface.co/datasets/allenai/reward-bench-results/blob/main/best-of-n/alpaca_eval/tulu-13b/OpenAssistant/oasst-rm-2.1-pythia-1.4b-epoch-2.5.json\r\n\r\nThere are more in that folder, ~40mb maybe?", "@albertvillanova here's a snippet so you don't need to click\r\n```\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 0\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 3.076171875\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 1\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 3.87890625\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 2\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 3.287109375\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 3\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 1.6337890625\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 4\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 5.27734375\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 5\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 3.0625\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 6\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 2.29296875\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 7\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 6.77734375\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 8\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 3.853515625\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 9\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 4.86328125\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 10\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 2.890625\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 11\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 4.70703125\r\n}\r\n{\r\n \"config\": \"top_p=0.9;temp=1.0\",\r\n \"dataset_details\": \"helpful_base\",\r\n \"id\": [\r\n 0,\r\n 12\r\n ],\r\n \"model\": \"allenai/tulu-2-dpo-13b\",\r\n \"scores\": 4.45703125\r\n}\r\n```", "Thanks again for your feedback, @natolambert.\r\n\r\nHowever, strictly speaking, the last file is not in JSON format but in kind of JSON-Lines like format (although not properly either because there are multiple newline characters within each object). Not even pandas can read that file format.\r\n\r\nAnyway, for JSON-Lines, I would expect that `datasets` and `pandas` have the same performance for JSON Lines files, as both use `pyarrow` under the hood...\r\n\r\nA proper JSON file in records orient should be a list (a JSON array): the first character should be `[`.\r\n\r\nAnyway, I am generating a JSON file from your JSON-Lines file to test performance." ]

[]

[]

[]

[]

[ { "login": "albertvillanova", "id": 8515462, "node_id": "MDQ6VXNlcjg1MTU0NjI=", "avatar_url": "https://avatars.githubusercontent.com/u/8515462?v=4", "gravatar_id": "", "url": "https://api.github.com/users/albertvillanova", "html_url": "https://github.com/albertvillanova", "followers_url": "https://api.github.com/users/albertvillanova/followers", "following_url": "https://api.github.com/users/albertvillanova/following{/other_user}", "gists_url": "https://api.github.com/users/albertvillanova/gists{/gist_id}", "starred_url": "https://api.github.com/users/albertvillanova/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/albertvillanova/subscriptions", "organizations_url": "https://api.github.com/users/albertvillanova/orgs", "repos_url": "https://api.github.com/users/albertvillanova/repos", "events_url": "https://api.github.com/users/albertvillanova/events{/privacy}", "received_events_url": "https://api.github.com/users/albertvillanova/received_events", "type": "User", "site_admin": false } ]

[]

[]

[]

[]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Same issue here." ]

[]

[ { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "This should help fixing this issue: https://github.com/huggingface/datasets/pull/6832", "I'm having a similar issue when using splices:\r\n<img width=\"947\" alt=\"image\" src=\"https://github.com/huggingface/datasets/assets/28941213/2153faac-e1fe-4b6d-a79b-30b2699407e8\">\r\n<img width=\"823\" alt=\"image\" src=\"https://github.com/huggingface/datasets/assets/28941213/80919eca-eb6c-407d-8070-52642fdcee54\">\r\n<img width=\"914\" alt=\"image\" src=\"https://github.com/huggingface/datasets/assets/28941213/5219c201-e22e-4536-acc3-a922677785ff\">\r\n\r\n\r\nIt seems to be downloading, loading, and generating splits using the entire dataset." ]

[]

[]

[]

[]

[]

[ "Thanks for reporting! I've opened a PR with a fix.", "Thanks, @mariosasko! Related question (although I guess this is a feature request): could we have some kind of exponential back-off for these retries? Here's my reasoning:\r\n- If a one-time accidental error happens, you should retry immediately and will succeed immediately.\r\n- If the Hub has a small outage on the order of minutes, you don't want to retry on the order of hours. \r\n- If the Hub has a prologned outage of several hours, we don't want to keep retrying on the order of minutes.\r\n\r\nThere actually already exists an implementation for (clipped) exponential backoff in the HuggingFace suite ([here](https://github.com/huggingface/huggingface_hub/blob/61b156a4f2e5fe1a492ed8712b26803e2122bde0/src/huggingface_hub/utils/_http.py#L306)), but I don't think it is used here.\r\n\r\nThe requirements are basically that you have an initial minimum waiting time and a maximum waiting time, and with each retry, the waiting time is doubled. We don't want to overload your servers with needless retries, especially when they're down :sweat_smile:", "Oh, I've just remembered that we added retries to the `HfFileSystem` in `huggingface_hub` 0.21.0 (see [this](https://github.com/huggingface/huggingface_hub/blob/61b156a4f2e5fe1a492ed8712b26803e2122bde0/src/huggingface_hub/hf_file_system.py#L703)), so I'll close the linked PR as we don't want to retry the retries :).\r\n\r\nI agree with the exponential backoff suggestion, so I'll open another PR.", "@mariosasko The call you linked indeed points to the implementation I linked in my previous comment, yes, but it has no configurability. Arguably, you want to have this hidden backoff under the hood that catches small network disturbances on the time scale of seconds -- perhaps even with hardcoded limits as is the case currently -- but you also still want to have a separate backoff on top of that with the configurability as suggested by @lhoestq in [the comment I linked](https://github.com/huggingface/datasets/issues/6172#issuecomment-1794876229).\r\n\r\nMy particular use-case is that I'm streaming a dataset while training on a university cluster with a very long scheduling queue. This means that when the backoff runs out of retries (which happens in under 30 seconds with the call you linked), I lose my spot on the cluster and have to queue for a whole day or more. Ideally, I should be able to specify that I want to retry for 2 to 3 hours but with more and more time between requests, so that I can smooth over hours-long outages without a setback of days.", "I also have my runs crash a surprising amount due to the dataloader crashing because of the hub, some way to address this would be nice." ]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[ "There are 2 workarounds, tho:\r\n1. Download datasets from web and just load them locally\r\n2. Use metadata directly (temporal solution, since metadata can change)\r\n```\r\nimport datasets\r\nfrom datasets.data_files import DataFilesDict, DataFilesList\r\n\r\ndata_files_list = DataFilesList(\r\n [\r\n \"hf://datasets/allenai/c4@1588ec454efa1a09f29cd18ddd04fe05fc8653a2/en/c4-train.00000-of-01024.json.gz\"\r\n ],\r\n [(\"allenai/c4\", \"1588ec454efa1a09f29cd18ddd04fe05fc8653a2\")],\r\n)\r\ndata_files = DataFilesDict({\"train\": data_files_list})\r\nc4_dataset = datasets.load_dataset(\r\n path=\"allenai/c4\",\r\n data_files=data_files,\r\n split=\"train\",\r\n cache_dir=\"/datesets/cache\",\r\n download_mode=\"reuse_cache_if_exists\",\r\n token=False,\r\n)\r\n```\r\nSecond solution also shows where to find the bug. I suggest that the hashing functions should always use only original parameter `data_files`, and not the one they get after connecting to the server and creating `DataFilesDict`", "Hi! You need to set the `HF_DATASETS_OFFLINE` env variable to `1` to load cached datasets offline, as explained in the docs [here](https://huggingface.co/docs/datasets/v2.19.0/en/loading#offline).", "Just tested. It doesn't work, because of the exact problem I described above: hash of dataset config is different.\r\nThe only error difference is the reason why it cannot connect to HuggingFace (now it's 'offline mode is enabled')\r\n\r\n" ]

[]

[]

[ "Get same error on same datasets too.", "+1", "same error" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6835). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.", "Fixed the conversion from `pyarrow` to `python` `Sequence` features. \r\n\r\nThere is still an issue that if `features` are passed the `Sequence` always forces conversion to `ListArray`.\r\nThis probably causes issues if the `LargeListArray` is actually needed.\r\n\r\nThere doesn't seem to be a great solution since this list is created solely on the `schema` for `Sequence`.\r\nOne solution would be to always use `LargeListArray` instead.\r\n" ]

[]

[]

[]

[]

[]

[ "Similar issue in text process \r\n\r\n```python\r\n\r\ntokenizer=AutoTokenizer.from_pretrained(model_dir[args.model])\r\ntrain_dataset=datasets.load_from_disk(dataset_dir[args.dataset],keep_in_memory=True)['train']\r\ntrain_dataset=train_dataset.map(partial(dname2func[args.dataset],tokenizer=tokenizer),batched=True,num_proc =50,remove_columns=train_dataset.features.keys(),desc='tokenize',keep_in_memory=True)\r\n\r\n```\r\nAfter this train_dataset will be like\r\n```python\r\nDataset({\r\n features: ['input_ids', 'labels'],\r\n num_rows: 51760\r\n})\r\n```\r\nIn which input_ids and labels are both List[int]\r\nHowever, per iter on dataset cost 7.412479639053345s ……？\r\n```python\r\nfor j in tqdm(range(len(train_dataset)),desc='first stage'):\r\n input_id,label=train_dataset['input_ids'][j],train_dataset['labels'][j]\r\n\r\n``` ", "The transform currently replaces the numpy formatting.\r\n\r\nSo you're back to copying data to long python lists which is super slow.\r\n\r\nIt would be cool for the transform to not remove the formatting in this case, but this requires a few changes in the lib" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6832). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

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

[]

[]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6828). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

[]

[]

[]

[]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[ "Hi ! You can try to reduce `writer_batch_size`. It corresponds to the number of samples that stay in RAM before being flushed to disk" ]

[]

[]

[ "This seems to be hardcoded behavior in table.py `array_cast`.\r\n```python\r\nif (\r\n not allow_number_to_str\r\n and pa.types.is_string(pa_type)\r\n and (pa.types.is_floating(array.type) or pa.types.is_integer(array.type))\r\n ):\r\n raise TypeError(\r\n f\"Couldn't cast array of type {array.type} to {pa_type} since allow_number_to_str is set to {allow_number_to_str}\"\r\n )\r\n if pa.types.is_null(pa_type) and not pa.types.is_null(array.type):\r\n raise TypeError(f\"Couldn't cast array of type {array.type} to {pa_type}\")\r\n return array.cast(pa_type)\r\n```\r\nwhere floats and integers are not cast to string but booleans are.\r\nMaybe this should be extended to booleans?", "Thanks for reporting! @Modexus Do you want to open a PR with the suggested fix?", "I'll gladly create a PR but not sure what the behavior should be.\r\n\r\nShould a value returned from map be cast to the current feature?\r\nAt the moment this seems very inconsistent since `datetime `is also cast (this would only fix `boolean`) but nested structures are not.\r\n\r\n```python\r\ndset = Dataset.from_dict({\"a\": [\"Hello world!\"]})\r\ndset = dset.map(lambda x: {\"a\": date(2021, 1, 1)})\r\n# dset[0][\"a\"] == '2021-01-01'\r\n```\r\n```python\r\ndset = Dataset.from_dict({\"a\": [\"Hello world!\"]})\r\ndset = dset.map(lambda x: {\"a\": [True]})\r\n# dset[0][\"a\"] == [True]\r\n```\r\n\r\nIs there are reason to cast the value if the user doesn't specify it explicitly?\r\nSeems tricky that some things are cast and some are not.", "Indeed, it also makes sense to raise a `TypeError` for temporal and decimal types.\r\n\r\n> Is there are reason to cast the value if the user doesn't specify it explicitly?\r\n\r\nThis is how PyArrow's built-in `cast` behaves - it allows casting from primitive types to strings. Hence, we need `allow_number_to_str` to disallow such casts (e.g., in the [scenario](https://github.com/huggingface/datasets/blob/a3bc89d8bfd47c2a175c3ce16d92b7307cdeafd6/src/datasets/arrow_writer.py#L208) when we are \"trying a type\" to preserve the original type if there is a column in the output dataset with the same name as in the input one).\r\n\r\nPS: In the PR, we can introduce `allow_numeric_to_str` (for floats, integers, decimals, booleans) and `allow_temporal_to_str` (for dates, timestamps, ...) and deprecate `allow_number_to_str` to make it clear what each parameter does.", "Would just `allow_primitive_to_str` work?\r\nThis should include all `numeric`, `boolean `and `temporal`formats.\r\n\r\nNote that at least in the [ C++ implementation](https://arrow.apache.org/docs/cpp/api/utilities.html#_CPPv410is_numericRK8DataType) `numeric `seems to exclude `boolean`.\r\n[](https://arrow.apache.org/docs/cpp/api/utilities.html#_CPPv410is_numericRK8DataType)", "Indeed, `allow_primitive_to_str` sounds better.\r\n\r\nPS: PyArrow's `pa.types.is_primitive` returns `False` for decimal types, but I think is okay for us to treat decimals as primitive types (or we can have `allow_decimal_to_str` to be fully consistent with PyArrow)" ]

[]

[]

[ "Hi !\r\n\r\nIt's possible to multiple files at once:\r\n\r\n```python\r\ndata_files = \"data/*.jsonl\"\r\n# Or pass a list of files\r\nlangs = ['ka-ml', 'br-sr', 'ka-pt', 'id-ko', ..., 'fi-ze_zh', 'he-kk', 'ka-tr']\r\ndata_files = [f\"data/{lang}.jsonl\" for lang in langs]\r\nds = load_dataset(\"loicmagne/open-subtitles-250-bitext-mining\", data_files=data_files, split=\"train\")\r\n```\r\n\r\nAlso maybe you can add a subset called \"all\" for people that want to load all the data without having to list all the languages ?\r\n\r\n```yaml\r\n - config_name: all\r\n data_files: data/*.jsonl\r\n```\r\n", "Thanks for your reply, it is indeed much faster, however the result is a dataset where all the subsets are \"merged\" together, the language pair is lost:\r\n```\r\nDatasetDict({\r\n train: Dataset({\r\n features: ['sentence1', 'sentence2'],\r\n num_rows: 247809\r\n })\r\n})\r\n```\r\nI guess I could add a 'lang' feature for each row in the dataset, is there a better way to do it ?", "Hi @lhoestq over at https://github.com/embeddings-benchmark/mteb/issues/530 we have started examining these issues and would love to make a PR for datasets if we believe there is a way to improve the speed. As I assume you have a better overview than me @lhoestq, would you be interested in a PR, and might you have an idea about where we would start working on it?\r\n\r\nWe see a speed comparison of \r\n1. 15 minutes (for ~20% of the languages) when loaded using a for loop\r\n2. 17 minutes using the your suggestion\r\n3. ~30 seconds when using @loicmagne \"merged\" method.\r\n\r\nWorth mentioning is that solution 2 looses the language information.", "Can you retry using `datasets` 2.19 ? We improved a lot the speed of downloading datasets with tons of small files.\r\n\r\n```\r\npip install -U datasets\r\n```\r\n\r\nNow this takes 17sec on my side instead of the 17min minutes @loicmagne mentioned :)\r\n\r\n```python\r\n>>> %time ds = load_dataset(\"loicmagne/open-subtitles-250-bitext-mining\", data_files=\"data/*.jsonl\")\r\nDownloading readme: 100%|█████████████████████████████████| 13.7k/13.7k [00:00<00:00, 5.47MB/s]\r\nResolving data files: 100%|█████████████████████████████████| 250/250 [00:00<00:00, 612.51it/s]\r\nDownloading data: 100%|██████████████████████████████████| 250/250 [00:12<00:00, 19.68files/s]\r\nGenerating train split: 247809 examples [00:00, 1057071.08 examples/s]\r\nCPU times: user 4.95 s, sys: 3.1 s, total: 8.05 s\r\nWall time: 17.4 s\r\n```", "> Can you retry using `datasets` 2.19 ? We improved a lot the speed of downloading datasets with tons of small files.\r\n> \r\n> ```\r\n> pip install -U datasets\r\n> ```\r\n> \r\n> Now this takes 17sec on my side instead of the 17min minutes @loicmagne mentioned :)\r\n> \r\n> ```python\r\n> >>> %time ds = load_dataset(\"loicmagne/open-subtitles-250-bitext-mining\", data_files=\"data/*.jsonl\")\r\n> Downloading readme: 100%|█████████████████████████████████| 13.7k/13.7k [00:00<00:00, 5.47MB/s]\r\n> Resolving data files: 100%|█████████████████████████████████| 250/250 [00:00<00:00, 612.51it/s]\r\n> Downloading data: 100%|██████████████████████████████████| 250/250 [00:12<00:00, 19.68files/s]\r\n> Generating train split: 247809 examples [00:00, 1057071.08 examples/s]\r\n> CPU times: user 4.95 s, sys: 3.1 s, total: 8.05 s\r\n> Wall time: 17.4 s\r\n> ```\r\n\r\nI was actually just noticing that, I bumped from 2.18 to 2.19 and got a massive speedup, amazing!\r\n\r\nAbout the fact that subset names are lost when loading all files at once, currently my solution is to add a 'lang' feature to each rows, convert to polars and use:\r\n\r\n```python\r\nds_split = ds.to_polars().group_by('lang')\r\n```\r\n\r\nIt's fast so I think it's an acceptable solution, but is there a better way to do it ?", "It's the fastest way I think :)\r\n\r\nAlternatively you can download the dataset repository locally using [huggingface_hub](https://huggingface.co/docs/huggingface_hub/guides/download) (either via CLI or in python) and load the subsets one by one locally using a for loop as you were doing before (just pass the directory path to load_dataset instead of the dataset_id). " ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Default `writer_batch_size `is set to 1000 (see [map](https://huggingface.co/docs/datasets/v2.16.1/en/package_reference/main_classes#datasets.Dataset.map)).\r\nThe \"tmp1335llua\" is probably the temp file it creates while writing to disk.\r\nMaybe try lowering the `writer_batch_size`.\r\n\r\nFor multi-processing you should probably pass the `processor `as an argument (with e.g. partial) to the function or create it inside so that the sub-processes have access to it and maybe add `if __name__ == \"__main__\"` (not sure that's necessary?).\r\n", "Hi @Modexus,\r\n\r\nThank you very much for the help! Yep after playing around with map, I managed to get the parallel processing to work by implementing it like you suggested.\r\n\r\nRegarding the temp files, it seems like the temp files just keep growing in size as the map continues. Eventually, once map finishes, the temp files are deleted, but they are instead saved as cache .arrow files. These cache files are absolutely gigantic (~ 30-50x the size of the initial dataset!).\r\n\r\nAfter playing around with the `prepare_dataset()` function above, it seems this issue is caused by the following line in the function, where the log-Mel spectrogram of the audio is calculated:\r\n\r\n`# compute log-Mel input features from input audio array\r\n batch[\"input_features\"] = processor.feature_extractor(audio[\"array\"], \r\n sampling_rate=audio[\"sampling_rate\"]).input_features[0]\r\n`\r\n\r\nWhen I remove this line, the final cache files are approximately the same size as the initial dataset.\r\n\r\nCan I check whether this is expected behavior with the whisper feature extractor? I cant imagine the spectrograms are that large!\r\n\r\nThank you so much for the help!", "I'm having a similar issue with the spectrographs taking up an incredibly large amount of space. (i.e. 100GB for 3GB of audio). Is this really normal behavior?", "Upon taking a look at the hex contents of the mapped dataset files I found that the overwhelming majority of the data contained within them was duplicated junk similar to this. I'm not very familiar with the inner workings of AI but I have to assume this is an inefficient way of storing data at best and a bug at worst.\r\n\r\n" ]

[]

[]

[ "From my current understanding, this timeout is only used when we need to get the results.\r\n\r\nOne of:\r\n1. All tasks are done\r\n2. One worker died\r\n\r\nYour function should work fine and it's definitely a bug if it doesn't.", "When one of the `map`'s worker processes crashes, the linked code re-raises an error from the crash and returns it to the caller.\r\n\r\nIf your question is how to limit the time of long-running tasks/worker processes, such functionality doesn't exist in `datasets` (yet), which means you need to implement it yourself.\r\n\r\nE.g., you can implement it using the built-in `signal` module like this:\r\n```python\r\nimport time\r\nimport signal\r\nfrom contextlib import contextmanager\r\n\r\nfrom datasets import Dataset\r\n\r\n\r\n@contextmanager\r\ndef max_exec_time(t):\r\n def raise_timeout_handler(signum, frame):\r\n raise TimeoutError\r\n \r\n orig_handler = signal.getsignal(signal.SIGALRM)\r\n signal.signal(signal.SIGALRM, raise_timeout_handler)\r\n try:\r\n signal.alarm(t)\r\n yield\r\n finally:\r\n signal.alarm(0)\r\n signal.signal(signal.SIGALRM, orig_handler)\r\n\r\n\r\ndef worker(example, rank):\r\n try:\r\n with max_exec_time(20): # 20 sec execution limit\r\n if rank % 2 == 0:\r\n time.sleep(50) # simulate a long-running task\r\n example[\"a\"] = 100\r\n except TimeoutError:\r\n example[\"a\"] = None # Or return empty batches here in the \"batched\" mode\r\n return example\r\n\r\ndata = Dataset.from_list([{\"a\": 1}, {\"a\": 2}])\r\ndata = data.map(worker, num_proc=2, with_rank=True)\r\nprint(data[0])\r\n```", "> From my current understanding, this timeout is only used when we need to get the results.\r\n> \r\n> One of:\r\n> \r\n> 1. All tasks are done\r\n> 2. One worker died\r\n> \r\n> Your function should work fine and it's definitely a bug if it doesn't.\r\n\r\nthanks for responding! can you reproduce the stuck with the above example code?", "> When one of the `map`'s worker processes crashes, the linked code re-raises an error from the crash and returns it to the caller.\r\n> \r\n> If your question is how to limit the time of long-running tasks/worker processes, such functionality doesn't exist in `datasets` (yet), which means you need to implement it yourself.\r\n> \r\n> E.g., you can implement it using the built-in `signal` module like this:\r\n> \r\n> ```python\r\n> import time\r\n> import signal\r\n> from contextlib import contextmanager\r\n> \r\n> from datasets import Dataset\r\n> \r\n> \r\n> @contextmanager\r\n> def max_exec_time(t):\r\n> def raise_timeout_handler(signum, frame):\r\n> raise TimeoutError\r\n> \r\n> orig_handler = signal.getsignal(signal.SIGALRM)\r\n> signal.signal(signal.SIGALRM, raise_timeout_handler)\r\n> try:\r\n> signal.alarm(t)\r\n> yield\r\n> finally:\r\n> signal.alarm(0)\r\n> signal.signal(signal.SIGALRM, orig_handler)\r\n> \r\n> \r\n> def worker(example, rank):\r\n> try:\r\n> with max_exec_time(20): # 20 sec execution limit\r\n> if rank % 2 == 0:\r\n> time.sleep(50) # simulate a long-running task\r\n> example[\"a\"] = 100\r\n> except TimeoutError:\r\n> example[\"a\"] = None # Or return empty batches here in the \"batched\" mode\r\n> return example\r\n> \r\n> data = Dataset.from_list([{\"a\": 1}, {\"a\": 2}])\r\n> data = data.map(worker, num_proc=2, with_rank=True)\r\n> print(data[0])\r\n> ```\r\n\r\nthanks for responding! However, I don't think we should use `signal` in the context of multiprocessing since sometimes it will crash one process and raise the following error\r\nhttps://github.com/huggingface/datasets/blob/c3ddb1ef00334a6f973679a51e783905fbc9ef0b/src/datasets/utils/py_utils.py#L664", "> thanks for responding! However, I don't think we should use signal in the context of multiprocessing since sometimes it will crash one process and raise the following error\r\n\r\nThe above code has `try/except` to catch the error from the handler. Or do you get an error other than `TimeoutError`?", "> > thanks for responding! However, I don't think we should use signal in the context of multiprocessing since sometimes it will crash one process and raise the following error\r\n> \r\n> The above code has `try/except` to catch the error from the handler. Or do you get an error other than `TimeoutError`?\r\n\r\nyup, it will raise the RuntimeError: https://github.com/huggingface/datasets/blob/c3ddb1ef00334a6f973679a51e783905fbc9ef0b/src/datasets/utils/py_utils.py#L667C19-L670C22\r\n\r\n```\r\n raise RuntimeError(\r\n \"One of the subprocesses has abruptly died during map operation.\"\r\n \"To debug the error, disable multiprocessing.\"\r\n )\r\n```" ]

[]

[]

[ "The docs for this PR live [here](https://moon-ci-docs.huggingface.co/docs/datasets/pr_6786). All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update." ]

[]

[]

[ "This may be a solution that only changes `cast_storage` of `Image`.\r\nHowever, I'm not totally sure that the assumptions hold that are made about the `ListArray`.\r\n\r\n```python\r\nelif pa.types.is_list(storage.type):\r\n from .features import Array3DExtensionType\r\n\r\n def get_shapes(arr):\r\n shape = ()\r\n while isinstance(arr, pa.ListArray):\r\n len_curr = len(arr)\r\n arr = arr.flatten()\r\n len_new = len(arr)\r\n shape = shape + (len_new // len_curr,)\r\n return shape\r\n\r\n def get_dtypes(arr):\r\n dtype = storage.type\r\n while hasattr(dtype, \"value_type\"):\r\n dtype = dtype.value_type\r\n return dtype\r\n\r\n arrays = []\r\n for i, is_null in enumerate(storage.is_null()):\r\n if not is_null.as_py():\r\n storage_part = storage.take([i])\r\n shape = get_shapes(storage_part)\r\n dtype = get_dtypes(storage_part)\r\n\r\n extension_type = Array3DExtensionType(shape=shape, dtype=str(dtype))\r\n array = pa.ExtensionArray.from_storage(extension_type, storage_part)\r\n arrays.append(array.to_numpy().squeeze(0))\r\n else:\r\n arrays.append(None)\r\n\r\n bytes_array = pa.array(\r\n [encode_np_array(arr)[\"bytes\"] if arr is not None else None for arr in arrays],\r\n type=pa.binary(),\r\n )\r\n path_array = pa.array([None] * len(storage), type=pa.string())\r\n storage = pa.StructArray.from_arrays(\r\n [bytes_array, path_array], [\"bytes\", \"path\"], mask=bytes_array.is_null()\r\n )\r\n```\r\n(Edited): to handle nulls\r\n\r\nNotably this doesn't change anything about the passing through of data or other things, just in the `Image` class.\r\nSeems quite fast:\r\n```bash\r\nFri Apr 5 17:55:51 2024 restats\r\n\r\n 63818 function calls (61995 primitive calls) in 0.812 seconds\r\n\r\n Ordered by: cumulative time\r\n List reduced from 1051 to 20 due to restriction <20>\r\n\r\n ncalls tottime percall cumtime percall filename:lineno(function)\r\n 47/1 0.000 0.000 0.810 0.810 {built-in method builtins.exec}\r\n 2/1 0.000 0.000 0.810 0.810 <string>:1(<module>)\r\n 2/1 0.000 0.000 0.809 0.809 arrow_dataset.py:594(wrapper)\r\n 2/1 0.000 0.000 0.809 0.809 arrow_dataset.py:551(wrapper)\r\n 2/1 0.000 0.000 0.809 0.809 arrow_dataset.py:2916(map)\r\n 3 0.000 0.000 0.807 0.269 arrow_dataset.py:3277(_map_single)\r\n 1 0.000 0.000 0.760 0.760 arrow_writer.py:589(finalize)\r\n 1 0.000 0.000 0.760 0.760 arrow_writer.py:423(write_examples_on_file)\r\n 1 0.000 0.000 0.759 0.759 arrow_writer.py:527(write_batch)\r\n 1 0.001 0.001 0.754 0.754 arrow_writer.py:161(__arrow_array__)\r\n 2/1 0.000 0.000 0.719 0.719 table.py:1800(wrapper)\r\n 1 0.000 0.000 0.719 0.719 table.py:1950(cast_array_to_feature)\r\n 1 0.006 0.006 0.718 0.718 image.py:209(cast_storage)\r\n 1 0.000 0.000 0.451 0.451 image.py:361(encode_np_array)\r\n 1 0.000 0.000 0.444 0.444 image.py:343(image_to_bytes)\r\n 1 0.000 0.000 0.413 0.413 Image.py:2376(save)\r\n 1 0.000 0.000 0.413 0.413 PngImagePlugin.py:1233(_save)\r\n 1 0.000 0.000 0.413 0.413 ImageFile.py:517(_save)\r\n 1 0.000 0.000 0.413 0.413 ImageFile.py:545(_encode_tile)\r\n 397 0.409 0.001 0.409 0.001 {method 'encode' of 'ImagingEncoder' objects}\r\n```", "Also encounter this problem. Has been strugging with it for a long time...", "This actually applies to all arrays (numpy or tensors like in torch), not only from external files.\r\n```python\r\nimport numpy as np\r\nimport datasets\r\n\r\nds = datasets.Dataset.from_dict(\r\n {\"image\": [np.random.randint(0, 255, (2048, 2048, 3), dtype=np.uint8)]},\r\n features=datasets.Features({\"image\": datasets.Image(decode=True)}),\r\n)\r\nds.set_format(\"numpy\")\r\n\r\nds = ds.map(load_from_cache_file=False)\r\n```" ]

[]

[]

[ "Hello!\r\n\r\nThis is due to how pandas write numpy arrays to csv. [Source](https://stackoverflow.com/questions/54753179/to-csv-saves-np-array-as-string-instead-of-as-a-list)\r\nTo fix this, you can convert them to list yourselves.\r\n\r\n```python\r\ndf = ds.to_pandas()\r\ndf['int'] = df['int'].apply(lambda arr: list(arr))\r\ndf.to_csv(index=False, '../output/temp.csv')\r\n```\r\n\r\nI think it would be good if `datasets` would do the conversion itself, but it's a breaking change and I would wait for the greenlight from someone from HF." ]

[]

[]

[ "Hi! `metadata.jsonl` (or `metadata.csv`) is the only allowed name for the `imagefolder`'s metadata files.", "@mariosasko hey i tried with metadata.jsonl also and it still doesn't get the right columns", "@mariosasko it says metadata.csv not found\r\n<img width=\"1150\" alt=\"image\" src=\"https://github.com/huggingface/datasets/assets/81643693/3754980c-6185-4413-88fa-b499bcdd4195\">\r\n\r\ndataset = load_dataset('/dataset',metadata.csv) \r\n\r\n| workspace\r\n|| source code\r\n| dataset\r\n| |-- images\r\n| |-- metadata.csv\r\n| |-- metadata.jsonl\r\n| |-- padded_images\r\n\r\nExample of metadata.jsonl file\r\n{\"caption\": \"a drawing depicts a full shot of a black t-shirt with a triangular pattern on the front there is a white label on the left side of the triangle\", \"image\": \"images/212734.png\", \"gaussian_padded_image\": \"padded_images/p_212734.png\"}\r\n{\"caption\": \"an eye-level full shot of a large elephant and a baby elephant standing in a watering hole on the left side is a small elephant with its head turned to the right of dry land, trees, and bushes\", \"image\": \"images/212735.png\", \"gaussian_padded_image\": \"padded_images/p_212735.png\"}\r\n", "Loading more than one image per row with `imagefolder` is not supported currently. You can subscribe to https://github.com/huggingface/datasets/issues/5760 to see when it will be.\r\n\r\nInstead, you can load the dataset with `Dataset.from_generator`:\r\n```python\r\nimport json\r\nfrom datasets import Dataset, Value, Image, Features\r\n\r\ndef gen():\r\n with open(\"./dataset/metadata.jsonl\") as f:\r\n for line in f:\r\n line = json.loads(line)\r\n yield {\"caption\": line[\"caption\"], \"image\": os.path.join(\"./dataset\", line[\"image\"], \"gaussian_padded_image\": os.path.join(\"./dataset\", line[\"gaussian_padded_image\"]))}\r\n\r\nfeatures = Features({\"caption\": Value(\"string\"), \"image\": Image(), \"gaussian_padded_image\": Image()})\r\ndataset = Dataset.from_generator(gen, features=features)\r\n```\r\n(E.g., if you want to share this dataset on the Hub, you can call `dataset.push_to_hub(...)` afterward)", "hi Thanks for sharing this, Actually I was trying with a webdataset format of the data as well and it did'nt work. Could you share how i can create Dataset object from webdataset format of this data?" ]

[]

[]

[ "Same problem.", "Hi! You should be able to fix this by passing `remove_unused_columns=False` to the `transformers` `TrainingArguments` as explained in https://github.com/huggingface/peft/issues/1299.\r\n\r\n(I'm not familiar with Vertex AI, but I'd assume `remove_unused_columns` can be passed as a flag to the docker container) ", "I had the same problem, but I spent a whole day trying different combination with my own dataset with the example data set and found the reason: the example data is multi-turn conversation between human and assistant, so # Humman or # Assistant appear at least twice. If your own custom data only has single turn conversation, it might end up with the same error. What you can do is repeat your single turn conversation twice in your training data (keep the key 'text' the same) and maybe it works. I guess the reason is the specific way processing the data requires and counts multi-turn only (single turn will be discarded so it ends up with no training data), but since I am using Google Vertex AI, I don't have direct access to the underlying code so that was just my guess. ", "> Hi! You should be able to fix this by passing `remove_unused_columns=False` to the `transformers` `TrainingArguments` as explained in [huggingface/peft#1299](https://github.com/huggingface/peft/issues/1299).\r\n> \r\n> (I'm not familiar with Vertex AI, but I'd assume `remove_unused_columns` can be passed as a flag to the docker container)\r\n\r\n@mariosasko Thanks for the response and suggestion. \r\nWhen I set `remove_unused_columns` as `False` , I end up getting different error (will post the error soon). \r\nEither the Vertex-AI does not support `remove_unused_columns` or my dataset is completely wrong. \r\n\r\nThank you, \r\nKK", "> I had the same problem, but I spent a whole day trying different combination with my own dataset with the example data set and found the reason: the example data is multi-turn conversation between human and assistant, so # Humman or # Assistant appear at least twice. If your own custom data only has single turn conversation, it might end up with the same error. What you can do is repeat your single turn conversation twice in your training data (keep the key 'text' the same) and maybe it works. I guess the reason is the specific way processing the data requires and counts multi-turn only (single turn will be discarded so it ends up with no training data), but since I am using Google Vertex AI, I don't have direct access to the underlying code so that was just my guess.\r\n\r\n@cyberyu Thanks for your suggestions. \r\nI have tried the approach you suggested, copied the same conversation in each jsonl element so every jsonl item has 2 `HUMAN` and `ASSISTANT`. \r\nHowever in my case, the issue persists. I am gonna give few more tries, and post the results here. \r\nYou can find my dataset [here](https://huggingface.co/datasets/kk2491/test/tree/main) \r\n\r\nThank you, \r\nKK ", "> > I had the same problem, but I spent a whole day trying different combination with my own dataset with the example data set and found the reason: the example data is multi-turn conversation between human and assistant, so # Humman or # Assistant appear at least twice. If your own custom data only has single turn conversation, it might end up with the same error. What you can do is repeat your single turn conversation twice in your training data (keep the key 'text' the same) and maybe it works. I guess the reason is the specific way processing the data requires and counts multi-turn only (single turn will be discarded so it ends up with no training data), but since I am using Google Vertex AI, I don't have direct access to the underlying code so that was just my guess.\r\n> \r\n> @cyberyu Thanks for your suggestions. I have tried the approach you suggested, copied the same conversation in each jsonl element so every jsonl item has 2 `HUMAN` and `ASSISTANT`. However in my case, the issue persists. I am gonna give few more tries, and post the results here. You can find my dataset [here](https://huggingface.co/datasets/kk2491/test/tree/main)\r\n> \r\n> Thank you, KK\r\n\r\nI think another reason is your training sample length is too short. I saw a relevant report (https://discuss.huggingface.co/t/indexerror-invalid-key-16-is-out-of-bounds-for-size-0/14298/16) stating that the processing code might have a bug discarding sequence length short than max_seq_length, which is 512. Not sure the Vertex AI backend code has fixed that bug or not. So I tried to add some garbage content in your data, and extended the length longer than 512 for a single turn, and repeated twice. You can copy the following line as 5 repeated lines as your training data jsonl file of five samples (no eval or test needed, for speed up, set evaluation step to 5 and training step to 10,), and it will pass.\r\n\r\n{\"text\":\"### Human: You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You will handle customers queries and provide effective help message. Please provide response to 'Can Interplai software optimize routes for minimizing package handling and transfer times in distribution centers'? ### Assistant: Yes, Interplai software can optimize routes for distribution centers by streamlining package handling processes, minimizing transfer times between loading docks and storage areas, and optimizing warehouse layouts for efficient order fulfillment. ### Human: You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You are a helpful AI Assistant familiar with customer service. You will handle customers queries and provide effective help message. Please provide response to 'Can Interplai software optimize routes for minimizing package handling and transfer times in distribution centers'? ### Assistant: Yes, Interplai software can optimize routes for distribution centers by streamlining package handling processes, minimizing transfer times between loading docks and storage areas, and optimizing warehouse layouts for efficient order fulfillment.\"}\r\n", "@cyberyu **Thank you so much, You saved my day (+ so many days)**. \r\nI tried the example you provided above, and the training is successfully completed in Vertex-AI (through GUI). \r\nI never thought there would be constraints on the length of the samples and also on the number of turns. \r\nI will update my complete dataset and see update here once the training is completed. \r\n\r\nThank you, \r\nKK " ]

[]

[]

[ "I think this is due to the speed of reading a `png` image using pillow compared to a `jpg` image.\r\nNotably the same is true with `tiff`, it is even faster than `jpg` in my case." ]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[ { "id": 1935892871, "node_id": "MDU6TGFiZWwxOTM1ODkyODcx", "url": "https://api.github.com/repos/huggingface/datasets/labels/enhancement", "name": "enhancement", "color": "a2eeef", "default": true, "description": "New feature or request" } ]

[]

[]

[]

[]

[ "I also need this feature for [\"Cnam-LMSSC/vibravox \"](https://huggingface.co/datasets/Cnam-LMSSC/vibravox)\r\n\r\n\r\nEDIT: Upgrading to `2.19.0` fixed my problem thanks to [this PR](https://github.com/huggingface/datasets/pull/6754)" ]