MothMalone/data-preprocessing-automl-benchmarks

Data Preprocessing AutoML Benchmarks

This repository contains text classification datasets with known data quality issues for preprocessing research in AutoML.

Usage

Load a specific dataset configuration like this:

from datasets import load_dataset
# Example for loading the TREC dataset
dataset = load_dataset("MothMalone/data-preprocessing-automl-benchmarks", "trec")

Available Datasets

Below are the details for each dataset configuration available in this repository.

Of course. Here are the completed descriptions for your dataset card.

imdb

• Description: A large movie review dataset for binary sentiment classification, containing 25,000 highly polarized movie reviews for training and 25,000 for testing.
• Data Quality Issue: N/A
• Classes: 2
• Training Samples: 18750
• Validation Samples: 6250
• Test Samples: 25000

twenty_newsgroups

• Description: A collection of approximately 20,000 newsgroup documents, partitioned evenly across 20 different newsgroups, making it a classic benchmark for text classification.
• Data Quality Issue: N/A
• Classes: 20
• Training Samples: 8485
• Validation Samples: 2829
• Test Samples: 7532

banking77

• Description: A fine-grained dataset of 13,083 customer service queries from the banking domain, annotated with 77 distinct intents.
• Data Quality Issue: N/A
• Classes: 77
• Training Samples: 7502
• Validation Samples: 2501
• Test Samples: 3080

trec

• Description: The Text REtrieval Conference (TREC) question classification dataset, containing questions categorized by their answer type (e.g., Person, Location, Number).
• Data Quality Issue: N/A
• Classes: 6
• Training Samples: 4089
• Validation Samples: 1363
• Test Samples: 500

financial_phrasebank

• Description: A collection of sentences from English financial news, annotated for sentiment (positive, negative, or neutral) by financial experts.
• Data Quality Issue: N/A
• Classes: 3
• Training Samples: 1358
• Validation Samples: 453
• Test Samples: 453

MASSIVE

• Description: A multilingual dataset of 1 million utterances for intent classification and slot filling, covering 52 languages. The en-US configuration is used here.
• Data Quality Issue: N/A
• Classes: 60
• Training Samples: 11514
• Validation Samples: 2033
• Test Samples: 2974