license: apache-2.0
task_categories:

• text-generation
  language:
• en
  tags:
• ai-safety
• adversarial
• critique
• reasoning
• trading
• finance
• red-teaming
  size_categories:
• 10M<n<100M
  pretty_name: CritiqueBank-11M

CritiqueBank-11M

The largest open dataset for training adversarial AI critics

📖 Overview

CritiqueBank-11M is a large-scale dataset of 11.7 million adversarial critique examples designed to train AI systems that detect reasoning flaws in autonomous AI outputs. Each example pairs a rationale (an AI-generated trading signal) with a critique that identifies potential flaws, biases, or blind spots.

This dataset powers MiniCrit, an adversarial validation system that has demonstrated 35% reduction in flawed AI outputs across 38,000+ production validations.

🎯 Purpose

Modern AI systems increasingly make autonomous decisions, but lack robust mechanisms for self-critique. CritiqueBank-11M enables training specialized "devil's advocate" models that:

📊 Dataset Statistics

🗂️ Data Fields

🔍 Example

💡 Critique Styles

🚀 Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("wmaousley/CritiqueBank-11M")

# View an example
print(dataset["train"][0])

# Filter by critique style
adversarial = dataset["train"].filter(lambda x: x["style"] == "adversarial")
print(f"Adversarial examples: {len(adversarial):,}")

🏋️ Training Usage

CritiqueBank-11M is designed for fine-tuning language models as adversarial critics:

def format_for_training(example):
    return f"""<|im_start|>system
You are MiniCrit, an adversarial AI critic. Find flaws in trading rationales.
<|im_end|>
<|im_start|>user
Critique this rationale: {example['rationale']}
<|im_end|>
<|im_start|>assistant
{example['critique']}<|im_end|>"""

# Apply to dataset
formatted = dataset["train"].map(lambda x: {"text": format_for_training(x)})

📈 Downstream Performance

Models trained on CritiqueBank-11M achieve:

📚 Citation

@dataset{ousley2026critiquebank,
  author    = {Ousley, William Alexander},
  title     = {CritiqueBank-11M: A Large-Scale Dataset for Training Adversarial AI Validators},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.18159342},
  url       = {https://doi.org/10.5281/zenodo.18159342}
}

🏢 About Antagon

Antagon Inc. builds adversarial validation systems for autonomous AI. We make AI systems trustworthy by training specialized critics that catch reasoning flaws before they cause failures.

🙏 Acknowledgments

⚖️ License

Apache 2.0 — Free for commercial and research use with attribution.