Cache management

When you download a dataset, the processing scripts and data are stored locally on your computer. The cache allows 🤗 Datasets to avoid re-downloading or processing the entire dataset every time you use it.

This guide will show you how to:

• Change the cache directory.
• Control how a dataset is loaded from the cache.
• Clean up cache files in the directory.
• Enable or disable caching.

Cache directory

The default cache directory is ~/.cache/huggingface/datasets. Change the cache location by setting the shell environment variable, HF_DATASETS_CACHE to another directory:

$ export HF_DATASETS_CACHE="/path/to/another/directory"

When you load a dataset, you also have the option to change where the data is cached. Change the cache_dir parameter to the path you want:

>>> from datasets import load_dataset
>>> dataset = load_dataset('LOADING_SCRIPT', cache_dir="PATH/TO/MY/CACHE/DIR")

Similarly, you can change where a metric is cached with the cache_dir parameter:

>>> from datasets import load_metric
>>> metric = load_metric('glue', 'mrpc', cache_dir="MY/CACHE/DIRECTORY")

Download mode

After you download a dataset, control how it is loaded by datasets.load_dataset() with the download_mode parameter. By default, 🤗 Datasets will reuse a dataset if it exists. But if you need the original dataset without any processing functions applied, re-download the files as shown below:

>>> from datasets import load_dataset
>>> dataset = load_dataset('squad', download_mode='force_redownload')

Refer to datasets.DownloadMode for a full list of download modes.

Cache files

Clean up the cache files in the directory with datasets.Dataset.cleanup_cache_files():

# Returns the number of removed cache files
>>> dataset.cleanup_cache_files()
2

Enable or disable caching

If you’re using a cached file locally, it will automatically reload the dataset with any previous transforms you applied to the dataset. Disable this behavior by setting the argument load_from_cache=False in datasets.Dataset.map():

>>> updated_dataset = small_dataset.map(add_prefix, load_from_cache=False)

In the example above, 🤗 Datasets will execute the function add_prefix over the entire dataset again instead of loading the dataset from its previous state.

Disable caching on a global scale with datasets.disable_caching():

>>> from datasets import disable_caching
>>> disable_caching()

When you disable caching, 🤗 Datasets will no longer reload cached files when applying transforms to datasets. Any transform you apply on your dataset will be need to be reapplied.

You can also avoid caching your metric entirely, and keep it in CPU memory instead:

>>> from datasets import load_metric
>>> metric = load_metric('glue', 'mrpc', keep_in_memory=True)

Improve performance

Disabling the cache and copying the dataset in-memory will speed up dataset operations. There are two options for copying the dataset in-memory:

1. Set datasets.config.IN_MEMORY_MAX_SIZE to a nonzero value (in bytes) that fits in your RAM memory.

2. Set the environment variable HF_DATASETS_IN_MEMORY_MAX_SIZE to a nonzero value. Note that the first method takes higher precedence.