--- dataset_info: features: - name: text dtype: string - name: label dtype: class_label: names: '0': World '1': Sports '2': Business '3': Sci/Tech splits: - name: test num_bytes: 1879474 num_examples: 7600 - name: val_full num_bytes: 5963460.6 num_examples: 24000 - name: val1 num_bytes: 2981730.3 num_examples: 12000 - name: val2 num_bytes: 2981730.3 num_examples: 12000 - name: train_full num_bytes: 23853842.4 num_examples: 96000 - name: train_1_48k num_bytes: 11926921.2 num_examples: 48000 - name: train_1_24k num_bytes: 5963460.6 num_examples: 24000 - name: train_1_12k num_bytes: 2981730.3 num_examples: 12000 - name: train_1_6k num_bytes: 1490865.15 num_examples: 6000 - name: train_1_3k num_bytes: 745432.575 num_examples: 3000 - name: train_1_1.5k num_bytes: 372716.2875 num_examples: 1500 download_size: 41027925 dataset_size: 61141363.7125 configs: - config_name: default data_files: - split: test path: data/test-* - split: val_full path: data/val_full-* - split: val1 path: data/val1-* - split: val2 path: data/val2-* - split: train_full path: data/train_full-* - split: train_1_48k path: data/train_1_48k-* - split: train_1_24k path: data/train_1_24k-* - split: train_1_12k path: data/train_1_12k-* - split: train_1_6k path: data/train_1_6k-* - split: train_1_3k path: data/train_1_3k-* - split: train_1_1.5k path: data/train_1_1.5k-* pretty_name: "Stratified and Nested Subsets of AG News for Benchmarking" tags: - ag_news - text-classification - stratified-sampling - low-resource - benchmarking license: apache-2.0 --- ## Stratified and Nested Subsets of AG News for Performance Benchmarking ### Dataset Summary This repository contains a script to generate stratified and progressively smaller, nested subsets of the **AG News** dataset. It was specifically designed to **benchmark the performance** (e.g., accuracy, training time, and resource usage) of language models on varying amounts of training data. This new version improves the benchmarking process by: 1. **Preserving the Original Test Set**: It uses the original `ag_news` test set for final, unbiased evaluation. 2. **Creating Dedicated Validation Sets**: It carves out multiple validation sets (`val_full`, `val1`, `val2`) from the original training data, allowing for robust hyperparameter tuning without touching the test set. By using **stratified samples**, each subset maintains the original class distribution of the AG News dataset. This ensures that performance differences can be more reliably attributed to dataset size rather than class imbalance. The training splits are **nested**, meaning `train_1_1.5k` is a perfect subset of `train_1_3k`, which is a subset of `train_1_6k`, and so on. --- ### Supported Tasks and Leaderboards * **`text-classification`**: The dataset is ideal for training models for news topic classification. The four classes are: World (0), Sports (1), Business (2), and Sci/Tech (3). --- ### Dataset Structure The dataset generation script creates multiple splits organized into three categories: **training**, **validation**, and **testing**. * The **`test`** split is the original, untouched test set from the `ag_news` dataset. * The **`validation`** splits are derived from 20% of the original `ag_news` training data. * The **`training`** splits are progressively smaller, nested subsets derived from the remaining 80% of the original `ag_news` training data. | Category | Split Name | Number of Samples | Description | | :----------- | :------------- | :---------------- | :---------------------------------------------------------- | | **Test** | `test` | 7,600 | The original `ag_news` test set. Use for final evaluation. | | **Validation** | `val_full` | 24,000 | 20% of the original training data. | | | `val1` | 12,000 | A 50% split of `val_full`. | | | `val2` | 12,000 | The other 50% split of `val_full`. | | **Training** | `train_full` | 96,000 | 80% of the original training data (full training set). | | | `train_1_48k` | 48,000 | 1/2 of `train_full`. | | | `train_1_24k` | 24,000 | 1/4 of `train_full`. | | | `train_1_12k` | 12,000 | 1/8 of `train_full`. | | | `train_1_6k` | 6,000 | 1/16 of `train_full`. | | | `train_1_3k` | 3,000 | 1/32 of `train_full`. | | | `train_1_1.5k` | 1,500 | 1/64 of `train_full`. | #### Data Fields * `text`: a `string` feature containing the news article content. * `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: ```python from datasets import load_dataset repo_name = "szhuggingface/ag_news" dataset = load_dataset(repo_name) # You can now access any split print("Medium training set:") print(dataset['train_1_6k']) print("\nTest set:") print(dataset['test']) ``` Instead of loading pre-made splits from the Hub, this repository provides a script to generate the datasets locally. This ensures transparency and allows for easy modification of the splitting parameters (like `random_state`). Save the code below as a Python file (e.g., `create_splits.py`) and run it. ```python import pandas as pd from datasets import load_dataset, DatasetDict def create_stratified_splits( dataset_name="ag_news", label_col="label", validation_size=0.2, random_state=42 ): """ Loads the ag_news train and test splits. It creates multiple validation and training sets for benchmarking purposes. - The final 'test' set is the original ag_news test set. - Validation sets ('val_full', 'val1', 'val2') are derived from the original train set. - Training sets are nested, stratified subsets of the remaining training data. Args: dataset_name (str): The name of the dataset on the Hugging Face Hub. label_col (str): The name of the column to stratify by. validation_size (float): The proportion of the original train set to use for validation. random_state (int): The seed for random operations to ensure reproducibility. Returns: DatasetDict: A dictionary containing all the created dataset splits. """ print(f"▶️ Loading original '{dataset_name}' train and test splits...") original_train_ds = load_dataset(dataset_name, split='train') original_test_ds = load_dataset(dataset_name, split='test') # 1. Create the main Train/Validation stratified split from the ORIGINAL TRAIN data print(f"\nSplitting original training data into a main training set ({1-validation_size:.0%}) and a full validation set ({validation_size:.0%})...") main_splits = original_train_ds.train_test_split( test_size=validation_size, stratify_by_column=label_col, seed=random_state ) train_full = main_splits['train'] val_full = main_splits['test'] # 2. Split the full validation set into two halves print("Creating two validation splits (val1, val2) from the full validation set...") val_splits = val_full.train_test_split( test_size=0.5, stratify_by_column=label_col, seed=random_state ) val1 = val_splits['train'] val2 = val_splits['test'] # 3. Create nested, growing subsets from the main training data print("Creating nested training subsets...") train_1_48k = train_full.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] train_1_24k = train_1_48k.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] train_1_12k = train_1_24k.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] train_1_6k = train_1_12k.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] train_1_3k = train_1_6k.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] train_1_1_5k = train_1_3k.train_test_split(train_size=0.5, stratify_by_column=label_col, seed=random_state)['train'] # 4. Consolidate all splits into a single DatasetDict final_datasets = DatasetDict({ 'test': original_test_ds, 'val_full': val_full, 'val1': val1, 'val2': val2, 'train_full': train_full, 'train_1_48k': train_1_48k, 'train_1_24k': train_1_24k, 'train_1_12k': train_1_12k, 'train_1_6k': train_1_6k, 'train_1_3k': train_1_3k, 'train_1_1.5k': train_1_1_5k }) print("\n--- ✅ Final Dataset Statistics ---") for name, ds in final_datasets.items(): print(f"\nDataset Split: '{name}' | Total samples: {len(ds):,}") return final_datasets ``` ### Citation 📜 If you use this dataset in your work, please cite the original paper: ```bibtex @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} }