Data loader

How a training data loader works

A training data loader does five tasks between storage and the model.

Task	PyTorch DataLoader (docs v2.15)	TensorFlow tf.data
Read	Dataset, num_workers	Dataset.interleave
Change	Transforms in the dataset	Dataset.map
Set sequence	Sampler, shuffle=True	Dataset.shuffle
Group	batch_size, collate_fn	Dataset.batch
Give to the model	pin_memory=True	Dataset.prefetch

3 risks of a data loader

1. GPU wait time. Without prefetch, TensorFlow says that “your model is sitting idle” while the pipeline fetches data.

Study	Population	Result
Murray et al.	Millions of tf.data jobs, Google fleet, one month in 2020	20% of jobs spent more than a third of compute time in the input pipeline
Mohan et al.	Nine vision and audio models, four datasets	As much as 65% of epoch time in preprocessing
Mohan et al.	Same study, language models	Bert-Large and GNMT showed no data stalls

2. Damage to the model with no change to the data. Shumailov and colleagues attacked only batch sequence, so this differs from data poisoning. On CIFAR-10 with ResNet-18, test accuracy fell from 90.51% to 26.13%. This is a benchmark result, not a reported incident.

3. Unwanted network connections. A document loader that crawls web pages acts on links that it finds. In CVE-2026-26019, a crawled page could point one LangChain loader at cloud metadata services. This is a form of server side request forgery.

In practice

Four checks for security teams follow. They are our synthesis, and no source gives them as one list.

1. Examine dataset classes as code. In PyTorch, dataset access, file reads and transforms run in the worker process.
2. Show how batches were selected. Shumailov and colleagues say that this randomness “is rarely tested or enforced”. They tell developers to show its provenance.
3. Check the files. The Hugging Face load_dataset function has SHA256 checks of downloaded files, but they are off by default.
4. Update the loaders. Use @langchain/community 1.1.14 or a subsequent version. Hugging Face datasets 4.0.0 removed dataset scripts and trust_remote_code support.

Three facts can surprise a team:

• PyTorch shuffle is off by default.
• Each PyTorch worker gets a copy of an iterable dataset. The copies need different settings, or the workers give duplicate data.
• Mohan and colleagues found a need for 3 to 24 CPU cores for each GPU. Light models such as ResNet18 needed the most.

Where definitions disagree

“Data loader” has no single meaning. Examine the context before you compare two tools.

Term	Source	What it does	Output
DataLoader	PyTorch	Gives samples to training code	Batches
Input pipeline	TensorFlow	Reads and changes data for training	Dataset elements
Document loader	LangChain	Reads data from different sources	Document objects
Data connector, or Reader	LlamaIndex	Takes data from sources and formats	Document objects

A document loader is one part of an application such as retrieval-augmented generation. It does not make training batches.

Questions and answers

What does a data loader do in machine learning?

A data loader reads training samples from storage and changes them for the model. It sets their sequence and gives the model one batch at a time. The PyTorch DataLoader and the TensorFlow tf.data API are two examples.

What is the difference between a data loader and a document loader?

A training data loader gives batches of samples to a model during training. A document loader reads files or web pages into Document objects for an LLM application. LangChain uses the term document loader.

Can a data loader be a security risk?

Yes. In a benchmark, researchers cut model accuracy with only a change to batch sequence. Document loaders can also fetch addresses that an attacker chooses, as CVE-2026-26019 showed in a LangChain web loader.