HinTox — Hinglish Hate Speech & Abuse Detector

HinTox is a binary text classifier trained to detect hate speech and abusive language in Hinglish — the code-mixed blend of Hindi and English widely used across Indian social media, messaging platforms, and online communities. A core strength of HinTox is its robustness to deliberate misspellings, leetspeak, and character-level obfuscation — techniques commonly used to evade standard content moderation filters.

---

Model Description

Property	Details
Task	Binary Text Classification
Labels	offensive, not offensive
Language	Hinglish (Hindi-English code-mixed)
Script	Roman (transliterated Hindi + English)
Base Model	fine-tuned transformer
Domain	Social media, online communities, chat

---

Key Features

🛡️ Obfuscation-Robust Detection

Standard hate speech detectors fail when users intentionally disguise slurs using character substitutions. HinTox is trained on text that mirrors real-world evasion tactics:

• Leetspeak / character substitution — e.g., madarch0d, bh0sdi, g4ndu, r4ndi
• Vowel-to-number mapping — a→4, o→0, e→3, i→1
• Asterisk masking — e.g., ch*tiye, b*kl
• Deliberate misspellings — e.g., bich instead of standard spelling, haramkhor
• Mixed obfuscation strategies — combining multiple techniques in a single token

🌐 Hinglish-Native

Most hate speech datasets are monolingual. HinTox is purpose-built for the code-mixed reality of Indian internet language, where speakers fluidly mix Hindi vocabulary, English grammar, and Roman script in the same sentence.

⚖️ Balanced & Contextual

The model distinguishes abusive language from casual Hinglish conversation, including slang, informal expressions, and friendly banter that may superficially resemble offensive content but is not.

---

Intended Use

HinTox is designed for:

• Content moderation pipelines on social platforms, forums, and comment sections
• Online safety tooling for apps with significant Hinglish-speaking user bases
• Research on code-mixed hate speech and NLP for low-resource/mixed language settings
• Dataset augmentation and annotation assistance for related tasks

---

Training Data

The model was trained on a curated dataset of Hinglish text with the following characteristics:

• Labels: offensive / not offensive
• Obfuscation coverage: leetspeak, asterisk masking, vowel substitution, deliberate misspellings
• Offensive examples span hate speech, gendered abuse, casteist slurs, and general profanity in Hinglish
• Non-offensive examples include casual conversation, everyday Hinglish slang, and friendly banter to reduce false positives

Example Data Points

"madarch0d samajhta kya hai apne aap ko", offensive
"bh0sdi ke aukaat mein reh apni", offensive
"bhai aaj match dekhne chalein kya", not offensive
"yaar mast gaana hai yeh", not offensive
"ch*tiye tu toh pura idiot hai", offensive
"chill kar bhai life mein", not offensive

---

Citation

If you use HinTox in your research or product, please cite:

@misc{hintox2025,
  title     = {HinTox: Obfuscation-Robust Hate Speech Detection for Hinglish},
  year      = {2026},
  note      = {HuggingFace Model Hub},
  url       = {https://huggingface.co/Keshav0av/HinTox}
}

---

Contact

For questions, feedback, or collaboration, open an issue on the model repository or reach out via HuggingFace.