Implementation:Huggingface Datasets Split

Knowledge Sources	Huggingface Datasets HF Datasets Docs
Domains	Data_Engineering, NLP
Last Updated	2026-02-14 18:00 GMT

Overview

Concrete tool for defining standard dataset split names provided by the HuggingFace Datasets library.

Description

Split is an enum-like class that defines the standard dataset split names: TRAIN, TEST, VALIDATION, and ALL. Each constant is a NamedSplit instance (a string subclass). Custom split names can be created by calling Split("custom_name"), which returns a NamedSplit for any name or a NamedSplitAll for "all". Split objects are used throughout the library for split identification in dataset construction, loading, and publishing.

Usage

Use Split constants to reference standard splits in dataset builders, generators, and loading functions. Use Split("name") to create custom split names.

Code Reference

Source Location

• Repository: datasets
• File: src/datasets/splits.py
• Lines: 407-453

Signature

class Split:
    TRAIN = NamedSplit("train")
    TEST = NamedSplit("test")
    VALIDATION = NamedSplit("validation")
    ALL = NamedSplitAll()

    def __new__(cls, name):
        """Create a custom split with datasets.Split('custom_name')."""
        return NamedSplitAll() if name == "all" else NamedSplit(name)

Import

from datasets import Split

I/O Contract

Inputs

Name	Type	Required	Description
name	str	Yes (for custom splits)	Name of the custom split. Not needed when using predefined constants.

Outputs

Name	Type	Description
instance	NamedSplit or NamedSplitAll	A split identifier usable throughout the Datasets library.

Usage Examples

Basic Usage

from datasets import Split, Dataset

# Use predefined constants
ds = Dataset.from_dict({"text": ["Hello"]}, split=Split.TRAIN)

# Create custom split
custom_split = Split("my_custom_split")

# Use in load_dataset
from datasets import load_dataset
ds = load_dataset("squad", split=Split.VALIDATION)

Related Pages

Implements Principle

• Principle:Huggingface_Datasets_Split_Definition