yashbleh/HiFox_Dataset

HiFoX Dataset

Overview

The HiFoX Dataset is a specialized, human-verified, and LLM-assisted benchmark dataset designed for Faux-Hate detection in Hindi-English (Hinglish) code-mixed text. Faux-Hate refers to the complex convergence of fake narratives and hate speech, where fake narratives are used to incite hostility. In code-mixed multilingual settings such as Hinglish, linguistic ambiguity, transliteration inconsistencies, and lexical variation complicate the identification of hate-wired fake narratives.

The dataset contains a total of 12,238 instances and is divided into two constituent phases/tasks:

• Task A (Fake and Hate Classification): Binary classification of Fake vs. Non-Fake and Hate vs. Non-Hate content.
• Task B (Fine-grained Hate Analysis): Multi-class classification focusing on the specific targets and severity of the hateful instances.

Dataset Structure

The dataset consists of Train, Validation, and Test splits for both Task A and Task B, provided in CSV format.

Data Splits

Split	Instances
Train	8,565
Validation	1,837
Test	1,836
Total	12,238

File Descriptions and Features

Task A (Train_Task_A.csv, Val_Task_A.csv, Test_Task_A.csv)

Task A focuses on the dual binary classification of Fake narratives and Hate Speech.

• Id: Unique identifier for the text instance (Integer).
• Tweet: The Hinglish code-mixed text (String).
• Hate: Binary label indicating the presence of genuine hate speech (1 for Hate, 0 for Non-Hate).
• Fake: Binary label indicating the presence of a fake narrative (1 for Fake, 0 for Non-Fake).

Task B (Train_Task_B.csv, Val_Task_B.csv, Test_Task_B.csv)

Task B provides a fine-grained, hierarchical analysis of the hateful instances, focusing on the target and severity.

• Id: Unique identifier for the text instance (Integer).
• Tweet: The Hinglish code-mixed text (String).
• Hate: Binary label indicating the presence of genuine hate speech (1 for Hate, 0 for Non-Hate).
• Target: Categorical label indicating the entity towards which the hate is targeted. Classes are:
  • I: Individual
  • O: Organization
  • R: Religion
  • (Blank for non-hateful instances)
• Severity: Categorical label quantifying the extent of hate. Classes are:
  • L: Low
  • M: Medium
  • H: High
  • (Blank for non-hateful instances)

Usage

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

# Load Task A data
dataset_task_a = load_dataset("yashbleh/HiFox_Dataset", data_files={
    "train": "Train_Task_A.csv",
    "validation": "Val_Task_A.csv",
    "test": "Test_Task_A.csv"
})

# Load Task B data
dataset_task_b = load_dataset("yashbleh/HiFox_Dataset", data_files={
    "train": "Train_Task_B.csv",
    "validation": "Val_Task_B.csv",
    "test": "Test_Task_B.csv"
})

Ethical Considerations

Given the nature of the task, the dataset inherently contains toxic, offensive, and hateful language. It should be used strictly for research purposes aimed at mitigating online harms and improving content moderation systems.

License and Terms of Use

This dataset is made available under Fair Use for academic and research purposes. However, as our corresponding research paper has not yet been published, any use of this dataset for research that leads to a publication requires explicit prior permission from the authors.

Users must ensure that their application of the dataset complies with relevant ethical guidelines and does not propagate harm. Please contact the authors to request permission before publishing any research utilizing this dataset.