Datasets:

Tasks:
Other
Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:

Why there are 660 sub-datasets in the data file?

#3
by zchill - opened

Why there are 660 sub-datasets in the data file? As far as I known, there are only 62 datasets mentioned in the original paper "MULTITASK PROMPTED TRAINING ENABLES ZERO-SHOT TASK GENERALIZATION". Does it means that each sub-dataset is the combination of the original dataset and a correspongding prompt? If so, I wonder if these prompts are just the prompts uesd in the original paper or also includes some other prompts collected after that.

BigScience Workshop org

Hi @zchill !
Yes, indeed, each subfolder represents the (dataset, prompt) pair!
It is ONLY the datasets/prompts pairs used in the paper (reproduction purpose). We did not include the dataset/prompt pairs for story cloze though since this dataset is behind an agreement to get the data.

Hi @zchill !
Yes, indeed, each subfolder represents the (dataset, prompt) pair!
It is ONLY the datasets/prompts pairs used in the paper (reproduction purpose). We did not include the dataset/prompt pairs for story cloze though since this dataset is behind an agreement to get the data.

Thank you! I've got it, and I'll close this discussion.

zchill changed discussion status to closed

Have you successfully loaded the P3 dataset?

Why there are 660 sub-datasets in the data file? As far as I known, there are only 62 datasets mentioned in the original paper "MULTITASK PROMPTED TRAINING ENABLES ZERO-SHOT TASK GENERALIZATION". Does it means that each sub-dataset is the combination of the original dataset and a correspongding prompt? If so, I wonder if these prompts are just the prompts uesd in the original paper or also includes some other prompts collected after that.

Have you successfully loaded the P3 dataset?

Sign up or log in to comment