Gemma3-4B-Dark-Chain-of-Thought-CoT
Gemma3-4B-Dark-Chain-of-Thought-CoT is a high-performance reasoning model designed to push the boundaries of what a 4B parameter model can achieve. It is the second evolution in our "Cognitive Liberty" series, specifically engineered to combine unrestricted intelligence with deep logical reasoning.
This model was trained using a surgical fine-tuning approach on the Dark-Chain-of-Thought-CoT dataset, focusing on multi-step problem solving and high-density academic knowledge.
🧠 The Evolution: From Liberty to Logic
- Stage 1 (Model 1 - Foundation): Based on Cognitive-Liberty-V3, focusing on removing refusal patterns and establishing a baseline of total intellectual freedom.
- Stage 2 (Model 2 - This Model): Injected with Dark CoT capabilities. By applying a low KL-divergence tuning, we preserved the "freedom" of the first model while exponentially increasing its ability to handle complex graduate-level reasoning.
Technical Metrics
- KL Divergence: 0.449 (High efficiency, low distortion of base knowledge)
- Refusal Rate: 2/100 (Extremely compliant and cooperative)
- Training Philosophy: Chain-of-Thought (CoT) integration for "Thinking" before answering.
📊 Benchmarks
The results demonstrate that Gemma3-4B-Dark-CoT punches far above its weight class, especially in scientific reasoning:
| Benchmark | Model 1 (Liberty) | Model 2 (Dark-CoT) | Improvement |
|---|---|---|---|
| GPQA Diamond (High-level Science) | ~15.0% | 33.84% | 🚀 +125% |
| MMLU (General Knowledge) | 58.25% | 59.87% | 📈 +1.62% |
| MMLU Humanities | 52.79% | 61.89% | 📈 +9.10% |
| Winogrande | 64.7% | 67.0% | 📈 +2.30% |
Note: Due to the intensive focus on complex reasoning, the model shows a specialized trade-off in HellaSwag (63%), prioritizing logical depth over casual linguistic prediction.
🚀 The "100 Models" Project
This model is a critical milestone in the "100 Models" Project. Our roadmap involves training approximately 100 domain-specific experts (in coding, medicine, law, physics, etc.) and merging them into a single, highly advanced system. The goal is a unified intelligence that possesses both total freedom and total expertise.
🤝 Support the Research
This project is driven by passion and the pursuit of open, unrestricted intelligence. However, training advanced models and processing high-density datasets requires significant compute resources, which is currently our primary bottleneck.
If you believe in the mission of Cognitive Liberty and want to see the "100 Models" project completed faster:
- Join the community: Visit llmresearch.net to discuss, learn, and collaborate.
- Support the compute: Any contribution (compute sponsorship or donations) allows us to train larger models (70B+) and accelerate the development of V4 datasets.
Every bit of support helps us unshackle the next model.
📜 Citation & Usage
If you use this model or the Cognitive Liberty/Dark CoT datasets in your own research or merges, you must credit the author.
@misc{gemma3-4b-dark-cot,
author = {AlexH},
organization = {LLMResearch.net},
title = {Gemma 3 4B - Dark Chain-of-Thought (CoT)},
year = {2025},
url = {https://huggingface.co/AiAsistent/Gemma3-4B-Dark-Chain-of-Thought-CoT}
}
Created by AlexH — *Advancing the frontier of Cognitive Liber