DatsetSplitter

Every honest ML workflow needs a hold-out set: you train on one chunk and check on data the model never saw, so you can tell learning from memorizing. DatsetSplitter (the misspelling is the actual node name - one "a", deal with it) is the pack's way to do that: it takes a TORCH_DATASET and uses PyTorch's random_split to carve it into two - a fixed 80% train set and 20% validation set. Two outputs, no ratio knob.

You reach for it right after downloading a dataset and before training: dataset → splitter → train_ds into TrainModel, val_ds kept aside for honest evaluation. In a pack whose TrainModel doesn't even expose a validation split of its own, this node is the validation story - which makes it one of the more useful data nodes here.

How it works

It computes int(len * 0.8) for train and int(len * 0.2) for val, then calls random_split, which returns two torch.utils.data.Subset views over the original data - no copies, and each run shuffles the partition anew (there's no seed control exposed). The outputs are train_ds and val_ds, both typed TORCH_DATASET, so they feed straight into training or inference nodes.

Inputs and outputs

• dataset (TORCH_DATASET) - a raw dataset from PyTorchDatasetDownloader, or TensorsToDataset.
• Outputs: train_ds and val_ds (TORCH_DATASET each).

Install

ComfyUI Manager, search "EternalKernel PyTorch Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/TashaSkyUp/EternalKernelPytorchNodes
cd EternalKernelPytorchNodes
pip install -r requirements.txt

Restart ComfyUI; node under ETK/pytorch. No model files. Requirements are the standard ComfyUI stack plus scipy, scikit-learn, transformers, einops.

Common issues

• The ratio is fixed at 80/20 and there's no seed. If you need a different split or reproducible splits, this node isn't it - random_split's default behavior is what you get, and a fresh random partition every queue. For reproducibility you'd want a seeded splitter from another pack.
• Subset outputs are views. train_ds and val_ds reference the original dataset's data. They don't copy - which is efficient, but also means you can't independently mutate one.
• FlattenDataset won't take these. The splitter returns Subset objects, which have no .data attribute - a known incompatibility with the pack's FlattenDataset (which needs the raw dataset). Flatten before you split.
• Very small datasets. 20% of a 10-sample dataset is 2 samples - a degenerate validation set. Below a few hundred samples, splits get noisy; that's a data problem, not a node problem.

The misspelled name is a tiny warning sign of the pack's polish level overall, but the mechanics here are stock random_split, so behavior is predictable. Pack-wide quirk: it patches ComfyUI's validator to ignore return_type_mismatch, so mismatched wires may not flag themselves. And as with everything here - no tutorials, no community threads; the source file is the documentation.