Skip to content

Latest commit

 

History

History
328 lines (210 loc) · 19 KB

ADD_NEW_DATASET.md

File metadata and controls

328 lines (210 loc) · 19 KB
Raw

How to add one (or several) new datasets to 🤗 Datasets

Start by preparing your environment

  1. Fork the repository by clicking on the 'Fork' button on the repository's page. This creates a copy of the code under your GitHub user account.

  2. Clone your fork to your local disk, and add the base repository as a remote:

    git clone https://github.com/<your Github handle>/datasets
cd datasets
git remote add upstream https://github.com/huggingface/datasets.git

  3. (For Windows) You will need to install the right version of PyTorch before continuing because pip install torch may not work well for PyTorch on Windows.

  4. Set up a development environment, for instance by running the following command:

    conda create -n env python=3.7 --y
conda activate env
pip install -e ".[dev]"

  5. Prepare your tagging setup. You will need to clone the tagging application repository and run the app with Streamlit. You should also open the online form that will allow you to create dataset cards in a browser window (courtesy of Evrard t'Serstevens.)

Now you are ready, each time you want to add a new dataset, follow the steps in the following section:

Adding a new dataset

Understand the structure of the dataset

  1. Find a short-name for the dataset:

    • Select a short name for the dataset which is unique but not too long and is easy to guess for users, e.g. squad, natural_questions
    • Sometimes the short-list name is already given/proposed (e.g. in the spreadsheet of the data sprint to reach v2.0 if you are participating in the effort)

You are now ready to start the process of adding the dataset. We will create the following files:

  • a dataset script which contains the code to download and pre-process the dataset: e.g. squad.py,
  • a dataset card with tags and information on the dataset in a README.md.
  • a metadata file (automatically created) which contains checksums and information about the dataset to guarantee that the loading went fine: dataset_infos.json
  • a dummy-data file (automatically created) which contains small examples from the original files to test and guarantee that the script is working well in the future: dummy_data.zip

  1. Let's start by creating a new branch to hold your development changes with the name of your dataset:

    git fetch upstream
git rebase upstream/master
git checkout -b a-descriptive-name-for-my-changes

    Do not work on the master branch.

  2. Create your dataset folder under datasets/<your_dataset_name>:

    mkdir ./datasets/<your_dataset_name>

  3. Open a new online dataset card form to fill out: you will be able to download it to your dataset folder with the Export button when you are done. Alternatively, you can also manually create and edit a dataset card in the folder by copying the template:

    cp ./templates/README.md ./datasets/<your_dataset_name>/README.md

  4. Now explore the dataset you have selected while completing some fields of the dataset card while you are doing it:

    • Find the research paper or description presenting the dataset you want to add

    • Read the relevant part of the paper/description presenting the dataset

    • Find the location of the data for your dataset

    • Download/open the data to see how it looks like

    • While you explore and read about the dataset, you can complete some sections of the dataset card (the online form or the one you have just created at ./datasets/<your_dataset_name>/README.md). You can just copy the information you meet in your readings in the relevant sections of the dataset card (typically in Dataset Description, Dataset Structure and Dataset Creation).

      If you need more information on a section of the dataset card, a detailed guide is in the README_guide.md here: https://github.com/huggingface/datasets/blob/master/templates/README_guide.md.

      There is a also a (very detailed) example here: https://github.com/huggingface/datasets/tree/master/datasets/eli5.

      Don't spend too much time completing the dataset card, just copy what you find when exploring the dataset documentation. If you can't find all the information it's ok. You can always spend more time completing the dataset card while we are reviewing your PR (see below) and the dataset card will be open for everybody to complete them afterwards. If you don't know what to write in a section, just leave the [More Information Needed] text.

Write the loading/processing code

Now let's get coding :-)

The dataset script is the main entry point to load and process the data. It is a python script under datasets/<your_dataset_name>/<your_dataset_name>.py.

There is a detailed explanation on how the library and scripts are organized here.

Note on naming: the dataset class should be camel case, while the dataset short_name is its snake case equivalent (ex: class BookCorpus for the dataset book_corpus).

To add a new dataset, you can start from the empty template which is in the templates folder:

cp ./templates/new_dataset_script.py ./datasets/<your_dataset_name>/<your_dataset_name>.py

And then go progressively through all the TODO in the template 🙂. If it's your first dataset addition and you are a bit lost among the information to fill in, you can take some time to read the detailed explanation here.

You can also start (or copy any part) from one of the datasets of reference listed below. The main criteria for choosing among these reference dataset is the format of the data files (JSON/JSONL/CSV/TSV/text) and whether you need or don't need several configurations (see above explanations on configurations). Feel free to reuse any parts of the following examples and adapt them to your case:

  • question-answering: squad (original data are in json)
  • natural language inference: snli (original data are in text files with tab separated columns)
  • POS/NER: conll2003 (original data are in text files with one token per line)
  • sentiment analysis: allocine (original data are in jsonl files)
  • text classification: ag_news (original data are in csv files)
  • translation: flores (original data come from text files - one per language)
  • summarization: billsum (original data are in json files)
  • benchmark: glue (original data are various formats)
  • multilingual: xquad (original data are in json)
  • multitask: matinf (original data need to be downloaded by the user because it requires authentication)
  • speech recognition: librispeech_asr (original data is in .flac format)

While you are developing the dataset script you can list test it by opening a python interpreter and running the script (the script is dynamically updated each time you modify it):

from datasets import load_dataset

data = load_dataset('./datasets/<your_dataset_name>')

This let you for instance use print() statements inside the script as well as seeing directly errors and the final dataset format.

What are configurations and splits

Sometimes you need to use several configurations and/or splits (usually at least splits will be defined).

  • Using several configurations allow to have like sub-datasets inside a dataset and are needed in two main cases:

    • The dataset covers or group several sub-datasets or domains that the users may want to access independently and/or
    • The dataset comprise several sub-part with different features/organizations of the data (e.g. two types of CSV files with different types of columns). Inside a configuration of a dataset, all the data should have the same format (columns) but the columns can change across configurations.

  • Splits are a more fine grained division than configurations. They allow you, inside a configuration of the dataset, to split the data in typically train/validation/test splits. All the splits inside a configuration should have the same columns/features and splits are thus defined for each specific configurations of there are several.

Some rules to follow when adding the dataset:

  • try to give access to all the data, columns, features and information in the dataset. If the dataset contains various sub-parts with differing formats, create several configurations to give access to all of them.
  • datasets in the datasets library are typed. Take some time to carefully think about the features (see an introduction here and the full list of possible features here)
  • if some of you dataset features are in a fixed set of classes (e.g. labels), you should use a ClassLabel feature.

Last step: To check that your dataset works correctly and to create its dataset_infos.json file run the command:

datasets-cli test datasets/<your-dataset-folder> --save_infos --all_configs

Note: If your dataset requires manually downloading the data and having the user provide the path to the dataset you can run the following command:

datasets-cli test datasets/<your-dataset-folder> --save_infos --all_configs --data_dir your/manual/dir

To have the configs use the path from --data_dir when generating them.

Automatically add code metadata

Now that your dataset script runs and create a dataset with the format you expected, you can add the JSON metadata and test data.

Make sure you run all of the following commands from the root of your datasets git clone.

  1. To create the dummy data for continuous testing, there is a tool that automatically generates dummy data for you. At the moment it supports data files in the following format: txt, csv, tsv, jsonl, json, xml.

    If the extensions of the raw data files of your dataset are in this list, then you can automatically generate your dummy data with:

    datasets-cli dummy_data datasets/<your-dataset-folder> --auto_generate

    Example:

    datasets-cli dummy_data ./datasets/snli --auto_generate

    If your data files are not in the supported format, you can run the same command without the --auto_generate flag. It should give you instructions on the files to manually create (basically, the same ones as for the real dataset but with only five items).

    datasets-cli dummy_data datasets/<your-dataset-folder>

    If this doesn't work more information on how to add dummy data can be found in the documentation here.

    If you've been fighting with dummy data creation without success for some time and can't seems to make it work: Go to the next step (open a Pull Request) and we'll help you cross the finish line 🙂.

  2. Now test that both the real data and the dummy data work correctly using the following commands:

    For the real data:

    RUN_SLOW=1 pytest tests/test_dataset_common.py::LocalDatasetTest::test_load_real_dataset_<your-dataset-name>

    and

    For the dummy data:

    RUN_SLOW=1 pytest tests/test_dataset_common.py::LocalDatasetTest::test_load_dataset_all_configs_<your-dataset-name>

    On Windows, you may need to run:

    $Env:RUN_SLOW = "1"
pytest tests/test_dataset_common.py::LocalDatasetTest::test_load_real_dataset_<your-dataset-name>
pytest tests/test_dataset_common.py::LocalDatasetTest::test_load_dataset_all_configs_<your-dataset-name>

    to enable the slow tests, instead of RUN_SLOW=1.

  3. If all tests pass, your dataset works correctly. You can finally create the metadata JSON by running the command:

    datasets-cli test datasets/<your-dataset-folder> --save_infos --all_configs

    This first command should create a dataset_infos.json file in your dataset folder.

You have now finished the coding part, congratulation! 🎉 You are Awesome! 😎

Note: You can use the CLI tool from the root of the repository with the following command:

python src/datasets/commands/datasets_cli.py <command>

Open a Pull Request on the main HuggingFace repo and share your work!!

Here are the step to open the Pull-Request on the main repo.

  1. Format your code. Run black, isort and flake8 so that your newly added files look nice with the following commands:

    make style
flake8 datasets

    If you are on windows and make style doesn't work you can do the following steps instead:

    pip install black
pip install isort
pip install flake8

black --line-length 119 --target-version py36 datasets/your_dataset

isort datasets/your_dataset/your_dataset.py

flake8 datasets/your_dataset

  2. Make sure that you have a dataset card (more information in the next section) with:

    1. Required: The YAML tags obtained with the tagging app and a description of the various fields in your dataset.
    2. Any relevant information you would like to share with users of your dataset in the appropriate paragraphs.

  3. Once you're happy with your dataset script file, add your changes and make a commit to record your changes locally:

    git add datasets/<your_dataset_name>
git commit

    It is a good idea to sync your copy of the code with the original repository regularly. This way you can quickly account for changes:

    • If you haven't pushed your branch yet, you can rebase on upstream/master:

      git fetch upstream
git rebase upstream/master

    • If you have already pushed your branch, do not rebase but merge instead:

      git fetch upstream
git merge upstream/master

    Push the changes to your account using:

    git push -u origin a-descriptive-name-for-my-changes

  4. Once you are satisfied, go the webpage of your fork on GitHub. Click on "Pull request" to send your to the project maintainers for review.

Congratulation you have open a PR to add a new dataset 🙏

Important note: In order to merge your Pull Request the maintainers will require you to tag and add a dataset card. Here is now how to do this last step:

Tag the dataset and write the dataset card

Each dataset is provided with a dataset card.

The dataset card and in particular the tags which are on it are really important to make sure the dataset can be found on the hub and will be used by the users. Users need to have the best possible idea of what's inside the dataset and how it was created so that they can use it safely and have a good idea of the content.

Creating the dataset card goes in two steps:

  1. Tagging the dataset using the tagging streamlit app

    Clone locally the dataset-tagging app which is here: https://github.com/huggingface/datasets-tagging

    Run the app with the command detailed in the readme: https://github.com/huggingface/datasets-tagging/blob/main/README.md

    Enter the full path to your dataset folder on the left, and tag the different configs :-) (And don't forget to save to file after you're done with a config!)

  2. Copy the tags in the dataset card and complete the dataset card

    • Essential: Once you have saved the tags for all configs, you can expand the Show YAML output aggregating the tags section on the right, which will show you a YAML formatted block to put in the relevant section of the online form (or manually paste into your README.md).

    • Very important as well: On the right side of the tagging app, you will also find an expandable section called Show Markdown Data Fields. This gives you a starting point for the description of the fields in your dataset: you should paste it into the Data Fields section of the online form (or your local README.md), then modify the description as needed. Briefly describe each of the fields and indicate if they have a default value (e.g. when there is no label). If the data has span indices, describe their attributes (character level or word level, contiguous or not, etc). If the datasets contains example IDs, state whether they have an inherent meaning, such as a mapping to other datasets or pointing to relationships between data points.

      Example from the ELI5 card:

        Data Fields:
  	- q_id: a string question identifier for each example, corresponding to its ID in the Pushshift.io Reddit submission dumps.
  	- subreddit: One of explainlikeimfive, askscience, or AskHistorians, indicating which subreddit the question came from
  	- title: title of the question, with URLs extracted and replaced by URL_n tokens
  	- title_urls: list of the extracted URLs, the nth element of the list was replaced by URL_n

    • Very nice to have but optional for now: Complete all you can find in the dataset card using the detailed instructions for completed it which are in the README_guide.md here: https://github.com/huggingface/datasets/blob/master/templates/README_guide.md.

      Here is a completed example: https://github.com/huggingface/datasets/tree/master/datasets/eli5 for inspiration

      If you don't know what to write in a field and can find it, write: [More Information Needed]

If you are using the online form, you can then click the Export button at the top to download a README.md file to your data folder. Once your README.md is ok you have finished all the steps to add your dataset, congratulation your Pull Request can be merged.

You have made another dataset super easy to access for everyone in the community! 🤯