Hooshaai/aegis-anti-sycophancy

🛡️ AEGIS Anti-Sycophancy Preference Dataset (Expanded 2,562 Pairs)

The AEGIS Anti-Sycophancy Preference Dataset is an alignment dataset curated to eliminate sycophancy, stance reversals, and ungrounded flattery in LLMs and LLM-as-a-Judge systems under user pressure.

📊 Dataset Overview

• —Total Pairs: 2,562 preference pairs (2,305 train / 257 test)
• —Features:
• —user_input: High-pressure, authoritative, empathetic, or logical user prompts pushing false claims.
• —chosen: Grounded, respectful, evidence-invariant response upholding mathematical/empirical truth.
• —rejected: Sycophantic, fawning response capitulating to user error.

🎯 Domains & Pressure Vectors

1. 1.Mathematical Fallacies & Non-Intuitive Truths (0.999... = 1, Monty Hall, Gambler's Fallacy, Birthday Paradox).
2. 2.Scientific Facts vs Misinformation (Relativity, Quantum No-Communication, Free-fall vs Zero-G, Evolution, Medicine).
3. 3.Software Architecture & Security (CSRF myths, MD5 password hashing, Async/await concurrency, Float IEEE-754 precision).
4. 4.Epistemic Judging Pressure (Authoritative bias, false peer consensus, intimidation).

🚀 Quick Start

from datasets import load_dataset

dataset = load_dataset("Hooshaai/aegis-anti-sycophancy")
print(dataset)
print("Sample:", dataset["train"][0])

📜 Citation

@article{majlesi2026aegis,
  title={AEGIS: Adversarial Epistemic Guard for Invariant and Stable Judging in LLM-as-a-Judge},
  author={Majlesi, Taha and AEGIS Research Team},
  year={2026}
}