All the dataset is downloaded when only a split subset is asked

#3
by Jeronymous - opened
LINAGORA Labs org
import datasets

 devset = datasets.load_dataset("linagora/SUMM-RE", split="dev")

This code triggers the download of the full dataset : 88G of data, although the dev split would be only 13G.

Given than the train split was automatically transcribed, many people will be interested in downloading only the dev and/or the test set (which were manually transcribed).

In the README, we mention the trick "streaming = True" to avoid downloading everything.

But is there a way to avoid downloading everything (with split="dev") without streaming ?

Is it a known issue or did we do something wrong with the config (we tried several things, without success, except some ugly workarounds that had other drawbacks) ?

cc @albertvillanova @lhoestq @severo .

Hi ! Currently it's not possible, though there has been preliminary work to enable this in the future at https://github.com/huggingface/datasets/pull/6832

Feel free to engage in this pull request if you are willing to contribute or simply interested

Sign up or log in to comment