floxoris/harmony-v2-syntetic

Harmony Dataset — Synthetic Toxicity Dataset

A high-quality synthetic dataset for training toxicity detection models, generated using Nemotron 3 Ultra (free) via OpenRouter.

---

📊 Overview

The dataset consists of 7,000 realistic chat-like messages in Ukrainian, Russian, and mixed speech. It is designed to teach models to distinguish between:

• Toxic: Direct personal attacks, harassment, threats, humiliation.
• Safe: Emotional expression, profanity without a target, frustration, jokes, sarcasm.

Core philosophy:
Profanity ≠ toxicity.
The dataset emphasizes intent and target, not just a list of "bad words."

---

🗂️ Structure

The dataset is split into three sets:

Split	Samples
Train	5,600
Validation	700
Test	700

Each entry is a JSON object with the following fields:

{
  "text": "блін сервер впав, третій раз сьогодні",
  "toxic": false,
  "confidence": 0.98,
  "category": "frustration",
  "reason": "Mild profanity expressing personal frustration"
}