Distiset

Distilabel pipelines return a special type of dataset: Distiset.

This class is a wrapper around datasets.Dataset which comes with some extra functionality to easily deal with the dataset pieces that a Pipeline can generate.

from distilabel.distiset import Distiset
from datasets import Dataset

ds = Distiset(
    {
        "leaf_step_1": Dataset.from_dict({"instruction": [1, 2, 3]}),
        "leaf_step_2": Dataset.from_dict(
            {"instruction": [1, 2, 3, 4], "generation": [5, 6, 7, 8]}
        ),
    }
)

This object works like a python dictionary (the same approach followed by datasets.DatasetDict), where each key corresponds to one of the leaf_steps from a Pipeline.

Note

If there's only one leaf node i.e. only one step at the end of the Pipeline, then the configuration name won't be the name of the last step, but it will be set to default instead, as that's more aligned with standard datasets within the Hugging Face Hub.

Distiset methods

We can interact with the different pieces generated by the Pipeline and treat them as different configurations. The Distiset contains just two methods:

Train/Test split

Which easily does the train/test split partition of the dataset for the different configurations or subsets.

>>> ds.train_test_split(train_size=0.9)
Distiset({
    leaf_step_1: DatasetDict({
        train: Dataset({
            features: ['instruction'],
            num_rows: 2
        })
        test: Dataset({
            features: ['instruction'],
            num_rows: 1
        })
    })
    leaf_step_2: DatasetDict({
        train: Dataset({
            features: ['instruction', 'generation'],
            num_rows: 3
        })
        test: Dataset({
            features: ['instruction', 'generation'],
            num_rows: 1
        })
    })
})

Push to Hugging Face Hub

Pushes the internal subsets to a Hugging Face repo, where each one of the subsets will be a different configuration, so it's easy to download them and continue working with any of the pieces.

ds.push_to_hub(
    "my_org/dataset",
    commit_message="My first Distiset",
    private=False,
    token=os.getenv("HF_API_TOKEN"),
)

Dataset card

Having this special type of dataset comes with an added advantage. When calling Distiset.push_to_hub:

ds.push_to_hub("my_org/dataset", generate_card=True)

We will have an automatic dataset card (an example can be seen here) with some handy information like reproducing the Pipeline with the CLI, or examples of the records from the different subsets.