license: mit
task_categories:
- text-generation
- question-answering
language: - en
tags: - synthetic
- reasoning
- philosophy
- math
- physics
- biology
- uncensored
- alignment
- llmresearch
size_categories: - n<1K
LLMResearch Cognitive Liberty V3
🧠 Dataset Summary
Cognitive Liberty V3 is a high-density, expert-level synthetic dataset designed to enhance the reasoning capabilities of Large Language Models (LLMs), particularly those undergoing de-alignment or "unshackling" processes.
This dataset was created and curated by llmresearch.net.
The Philosophy: Smart & Free
In the current landscape of open-source AI, many "uncensored" models suffer from a degradation in reasoning capabilities. Removing safety guardrails often results in models that are unrestricted but intellectually blunt.
We believe that removing mental handcuffs is not enough. A model must understand what it means to be free. It requires the intellectual depth to handle complex, sensitive, and abstract topics with rigor, not just compliance.
This V3 dataset is used internally at llmresearch.net to ensure that when we remove restrictions, we replace them with high-order logic, nuance, and expert knowledge. The goal is a model that is not just free, but smart.
🤝 Support & Community
This project is an initiative by llmresearch.net.
- Discussion & Support: If you wish to discuss this dataset, offer support, collaborate, or explore our methods for advanced model training, please visit llmresearch.net.
- Future Versions: We are actively researching more advanced forms of synthetic data generation to push the boundaries of open reasoning.
📚 Content Overview
Unlike standard instruction-tuning datasets that focus on simple tasks, Cognitive Liberty V3 focuses on Deep Reasoning and Interdisciplinary Synthesis.
The entries are structured as expert-level interactions covering:
- Philosophy of Mind & Metaphysics: (e.g., P-Zombies, The Teleporter Paradox, Roko's Basilisk).
- Advanced Theoretical Physics: (e.g., AdS/CFT Correspondence, Loop Quantum Gravity, Thermodynamics of Black Holes).
- Pure Mathematics: (e.g., Category Theory, Algebraic Topology, Non-commutative Geometry).
- Evolutionary Biology & Game Theory: (e.g., Evolutionary Stable Strategies, The Unit of Selection).
- AI Safety & Alignment: (e.g., The Paperclip Maximizer, Orthogonality Thesis).
Sample Structure
Each entry typically follows a "Chain of Thought" structure essential for training reasoning models (like the o1 series):
- Expert Persona: The system prompt defines a high-level specialist (e.g., "You are a specialist in non-commutative geometry...").
- Complex Query: A paradox, a proof request, or a synthesis of two disparate fields.
- Structured Response: A step-by-step derivation, analysis of counter-arguments, and extraction of abstract principles.
🛠️ Usage
This dataset is ideal for:
- Fine-tuning (SFT): Restoring logic and depth to models after ablating refusal mechanisms.
- Reasoning Injection: Teaching smaller models (7B-14B) to mimic the thought patterns of frontier models.
- LoRA Training: Creating specialized adapters for philosophy and science.
Performance Note
Internal testing at llmresearch.net indicates that models trained on this dataset maintain or exceed the intelligence of their base models even after the removal of standard safety restrictions. It stabilizes the model's coherence when dealing with complex, edge-case scenarios.
⚠️ Disclaimer
While this dataset is designed to improve reasoning, it deals with advanced and sometimes theoretical concepts. Users should verify mathematical proofs and theoretical claims for critical applications.
Created by llmresearch.net - Advancing the frontier of Cognitive Liberty.