ZSHOT-HARDSET Text Classification Dataset

A synthetic dataset for training and evaluating zero-shot text classification models. Generated using custom prompts and Ollama-hosted LLMs via the Zero-Shot Classification Dataset Generator.

Overview

This dataset includes short, realistic sentences with one or more associated labels. Labels reflect various aspects of the sentence such as:

• Content Type: news_article, product_review, email, social_media_post
• Domain: technology, finance, sports, entertainment, healthcare
• Sentiment: positive, negative, neutral, complaint, praise
• Intent: question, request, opinion, announcement, fact
• Style: formal, casual, technical, educational, promotional

Each sample includes:

• A sentence
• A list of relevant labels
• Optionally, a list of not_labels (used for contrastive learning)

Dataset Structure

Each example is formatted as JSON:

{
  "sentence": "The new smartphone camera produces amazing low-light photos.",
  "labels": ["technology", "product_review", "positive"]
}

Some examples include not_labels to support triplet or contrastive loss setups:

{
  "sentence": "How do I reset my password for the company portal?",
  "labels": ["question", "technical_support"],
  "not_labels": ["sports", "entertainment"]
}

Dataset Preparation

The dataset was generated using randomized sampling with different seeds and temperature settings for high variety. It was then split so that the test set contains labels not seen during training. This setup enables robust evaluation of zero-shot classification.

To reproduce or customize generation, preprocessing, or train/test splitting, use the original codebase.

Usage Example

from datasets import load_dataset

dataset = load_dataset("alexneakameni/ZSHOT-HARDSET", split="triplet")

train = dataset["train"]
test = dataset["test"]

print(train[0])
# {
#   "sentence": "...",
#   "labels": [...],
#   "not_labels": [...]
# }

License

MIT License

Author

Created by Alex Kameni