tasksource/num-glue

NumGLUE

This repository provides NumGLUE in a normalized Parquet layout for convenient loading from Hugging Face Datasets.

The default all config is balanced for unified training and evaluation: each NumGLUE type contributes at most 1,000 examples per split. It is ordered with prettyorder on type, using a 3-example interleaved prefix followed by a seeded shuffle (seed=42) of the remaining rows. The type_1 through type_8 configs expose each original NumGLUE type separately at full size with the same column schema as all.

Usage

from datasets import load_dataset

dataset = load_dataset("tasksource/num-glue")
type_5 = load_dataset("tasksource/num-glue", "type_5")

Columns

• id: deterministic row id built from split and source row index.
• type: original NumGLUE type label, from Type_1 to Type_8.
• question, passage, statement1, statement2, options, option1, option2: normalized input fields. Missing source fields are null.
• answer: simple string answer for direct use.
• answer_json: exact JSON serialization of the original answer value.
• answer_type: original JSON answer type (int, str, or dict).

Original NumGLUE_dev.json is exposed as the Hugging Face validation split.

Split Sizes

config	train	validation	test
all	6,502	3,492	4,042
type_1	282	41	81
type_2	1,131	164	325
type_3	564	81	162
type_4	770	110	220
type_5	37,949	5,421	4,835
type_6	22,908	3,272	3,015
type_7	6,791	970	1,691
type_8	886	126	254

Citation

Please cite the NumGLUE paper:

@article{mishra2022numglue,
  title = {NumGLUE: A Suite of Fundamental yet Challenging Mathematical Reasoning Tasks},
  author = {Mishra, Swaroop and Mitra, Arindam and Varshney, Neeraj and Sachdeva, Bhavdeep and Clark, Peter and Baral, Chitta and Kalyan, Ashwin},
  journal = {arXiv preprint arXiv:2204.05660},
  year = {2022},
  url = {https://arxiv.org/abs/2204.05660}
}