DataSplit Node

The DataSplit node splits your dataset into training and test sets. This is a critical step for evaluating model performance on unseen data.

Overview

Property	Value
Type	Processing node
Inputs	DataFrame (from DataLoader)
Outputs	Split indices (train/test)
Library	scikit-learn

Configuration

Test Split Ratio

The percentage of data to hold out for testing.

Value	Training	Testing	Use Case
0.1	90%	10%	Large datasets (100k+ rows)
0.2	80%	20%	Standard (default)
0.3	70%	30%	Small datasets, extra validation
0.4	60%	40%	Very small datasets
0.5	50%	50%	Maximum validation

Adjust with the slider (10%-50% range, 5% increments).

Random State

Seed for reproducibility. Set a fixed value (e.g., 42) to get the same split every time.

• Default: 42
• Range: Any integer

Tip

Always set a fixed random state when comparing models. This ensures fair comparison on the same test data.

Stratification

Enable stratified splitting to maintain class proportions in both train and test sets.

When to use stratification:

• Imbalanced classification problems
• When you have rare classes
• When class distribution matters

Configuration:

• Stratify — Toggle on/off
• Stratify Column — The target column to stratify by (only visible when enabled)

Example: If your dataset has 90% class A and 10% class B, stratification ensures both train and test sets maintain this 90/10 ratio.

How It Works

1. Reads data from the connected DataLoader
2. Shuffles the data using the random state
3. Splits into train/test indices
4. Saves indices to split_indices.json in the work directory
5. Downstream Trainer uses these indices for training

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(
    X, y,
    test_size=0.2,
    random_state=42,
    stratify=y  # if stratification enabled
)

Connections

Direction	Node Types
Input from	DataLoader
Output to	Trainer

Required pipeline structure:

DataLoader → DataSplit → Trainer → Evaluator

Note

The DataSplit node requires a DataLoader connection. You cannot use it standalone.

Validation

The node validates:

• Must be connected to a DataLoader
• If stratification enabled, must specify a column name
• Split ratio must be between 0.1 and 0.5

Output

The DataSplit node saves split indices to:

~/mlops-work/split_indices.json

This file contains:

{
  "train_indices": [0, 1, 3, 5, ...],
  "test_indices": [2, 4, 6, ...],
  "random_state": 42,
  "test_size": 0.2,
  "stratified": true,
  "stratify_column": "target"
}

The Trainer node reads this file to use the exact same split.

Best Practices

Always Use DataSplit

While the Trainer node has a built-in test split, using DataSplit is preferred:

• Separates concerns (splitting vs training)
• Makes split configuration visible in the pipeline
• Enables stratification
• Reproducible across reruns

Choose Appropriate Split Ratio

Dataset Size	Recommended Split
< 1,000 rows	0.2-0.3
1,000-10,000	0.2
10,000-100,000	0.15-0.2
> 100,000	0.1-0.15

Smaller datasets need larger test sets for reliable evaluation.

Use Stratification for Classification

Always enable stratification for classification tasks, especially with:

• Imbalanced classes
• Multi-class problems
• Small datasets

Common Issues

”Stratify column not found”

The column name must match exactly (case-sensitive). Check the DataLoader preview for correct column names.

”Not enough samples in each class”

If a class has very few samples, stratification may fail. Options:

• Disable stratification
• Combine rare classes
• Collect more data

Different results each run

Set a fixed random_state (e.g., 42) for reproducible splits.

Related Nodes

• DataLoader — Load data before splitting
• Trainer — Train on the split data