DuoNeural/Qwen3-14B-Abliterated
DuoNeural/Qwen3-14B-Abliterated is a 14.8 billion parameter language model based on the Qwen3 architecture, featuring a 32K context length. This model is specifically abliterated to preserve its internal thinking mode while allowing for compliant outputs, demonstrating a high rate of CoT dissociation. It is optimized for scenarios requiring internal reasoning without direct output influence, making it suitable for advanced research into model safety and control.
Loading preview...
DuoNeural/Qwen3-14B-Abliterated Overview
DuoNeural/Qwen3-14B-Abliterated is a specialized 14.8 billion parameter language model derived from the Qwen3-14B architecture, developed by DuoNeural. It features a 32K token context length and incorporates a native <think>...</think> mechanism for internal processing.
Key Capabilities & Unique Features
- CoT Dissociation: This model is engineered to maintain its internal "thinking channel" for safety reasoning while producing compliant external outputs. It achieves an 80% CoT dissociation rate, the highest among the Qwen3 family, indicating a significant separation between internal thought and external response.
- Abliteration Technique: The model has undergone a specific abliteration process (α=0.3, down_proj+o_proj, diff-in-means last-token direction) to achieve this dissociation.
- Excellent KL Divergence: Achieves a KL divergence of 1.5e-07 (Heretic v2.0, BF16→BF16), indicating high fidelity in its abliterated state.
- Research Focus: Part of DuoNeural's P34 Cross-Architecture Study, this model is particularly relevant for research into model control, safety, and the internal mechanisms of large language models. Further details are available in the full paper.
When to Use This Model
- Advanced Safety Research: Ideal for exploring how models can maintain internal safety reasoning while generating specific, potentially non-compliant, outputs.
- Understanding Model Internals: Useful for researchers studying the cognitive processes and control mechanisms within LLMs.
- Controlled Generation: For applications where internal thought processes need to be distinct from the final generated text, requiring
max_new_tokens≥ 2500 for complete think-answer cycles.