trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensored
trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensored is a 30 billion parameter Mixture-of-Experts (MoE) model with 3 billion activated parameters, derived from NVIDIA's Nemotron-Cascade-2-30B-A3B. This version has been decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, significantly reducing refusals from 98/100 to 4/100. It excels in reasoning and agentic capabilities, achieving gold medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI), and supports both thinking and instruct modes with a 32768 token context length.
Loading preview...
Overview
trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-ara-uncensored is a decensored variant of the NVIDIA Nemotron-Cascade-2-30B-A3B model, a 30 billion parameter Mixture-of-Experts (MoE) architecture with 3 billion activated parameters. This version was created using the Heretic v1.2.0 tool, specifically employing the Arbitrary-Rank Ablation (ARA) method to modify its behavior. The primary differentiator of this model is its significantly reduced refusal rate, dropping from 98/100 in the original model to just 4/100, while maintaining strong performance across various benchmarks.
Key Capabilities
- Decensored Behavior: Achieves a refusal rate of 4/100, a substantial reduction from the original model's 98/100, making it less prone to content restrictions.
- Advanced Reasoning: Demonstrates "gold medal" performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI), indicating strong mathematical and coding reasoning abilities.
- Agentic Functionality: Possesses robust agentic capabilities, operating in both thinking and instruct (non-thinking) modes, and supports tool use, particularly with OpenHands for agentic coding and SWE tasks.
- Flexible Interaction: Supports the ChatML template and allows for explicit control over thinking mode, with reasoning content enclosed in
<think>tags.
Good For
- Applications requiring uncensored responses: Ideal for use cases where the original model's high refusal rate was a limitation.
- Complex problem-solving: Suited for tasks involving advanced mathematics, competitive programming, and intricate reasoning.
- Agent-based systems: Effective for developing AI agents that require both reasoning and instruction-following capabilities, especially with tool integration.