szhuggingface/ag_news_test

Stratified and Nested Subsets of AG News for Performance Benchmarking

Dataset Summary

This repository contains stratified and progressively smaller, nested subsets of the AG News dataset. It was specifically created to benchmark the performance (e.g., accuracy, training time, and resource usage) of fine-tuning language models on varying amounts of training data.

By using stratified samples, each training subset maintains the original class distribution of the AG News dataset, ensuring that performance differences can be more reliably attributed to dataset size rather than class imbalance.

The splits are nested, meaning train_very_very_small is a perfect subset of train_very_small, which is a subset of train_small, and so on.

Supported Tasks and Leaderboards

• text-classification: The dataset can be used to train models for news topic classification. The four classes are: World (0), Sports (1), Business (2), and Sci/Tech (3).

Dataset Structure

The dataset contains six splits. A test split was created from 20% of the original data. The remaining 80% forms the train_full split, from which all other training sets were derived.

Split Name	Number of Samples	Description
test	24,000	20% of the original data, for evaluation.
train_full	96,000	80% of the original data (full training set).
train_medium	48,000	1/2 of the train_full set.
train_small	24,000	1/4 of the train_full set.
train_very_small	12,000	1/8 of the train_full set.
train_very_very_small	6,000	1/16 of the train_full set.

Data Fields

• text: a string feature.
• label: a classification label, with possible values including World (0), Sports (1), Business (2), Sci/Tech (3).

How to Use

You can load the dataset using the datasets library with a single line of code:

from datasets import load_dataset

repo_name = "szhuggingface/ag_news_test" 
dataset = load_dataset(repo_name)

# You can now access any split
print("Medium training set:")
print(dataset['train_medium'])

print("\nTest set:")
print(dataset['test'])

Citation 📜

If you use this dataset in your work, please cite the original paper:

@inproceedings{zhang2015character,
  title={Character-level convolutional networks for text classification},
  author={Zhang, Xiang and Zhao, Junbo and LeCun, Yann},
  booktitle={Advances in neural information processing systems},
  pages={649--657},
  year={2015}
}