saidutta69/DeepSeek-R1-Distill-Qwen-1.5B-heretic
DeepSeek-R1-Distill-Qwen-1.5B-heretic is a 1.5 billion parameter Qwen2.5-class language model developed by saidutta69, based on deepseek-ai's DeepSeek-R1-Distill-Qwen-1.5B. This variant is decensored using Heretic v1.4.0's directional ablation, which suppresses refusal behavior via targeted weight edits rather than fine-tuning. It is designed for developers seeking DeepSeek-R1's distilled reasoning capabilities without guardrail-induced refusals, running efficiently on consumer hardware and CPUs.
Loading preview...
Overview
This model, DeepSeek-R1-Distill-Qwen-1.5B-heretic, is a 1.5 billion parameter variant of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. Its primary distinction is the removal of refusal behaviors through a technique called "abliteration" (directional ablation) using Heretic v1.4.0. This method directly edits specific weight directions responsible for refusal, preserving the base model's knowledge, reasoning, and instruction-following capabilities without the degradation often seen with fine-tuning for helpfulness.
Key Capabilities & Features
- Decensored Reasoning: Provides the distilled reasoning of the DeepSeek-R1 base model without built-in refusal behaviors.
- Efficient Deployment: The 1.5B Qwen2.5-class core is designed to run on CPUs and consumer GPUs, with quantized versions fitting within 1-2 GB of VRAM.
- Abliteration Method: Utilizes targeted weight edits to suppress refusals, maintaining the original model's coherence and capabilities.
- GGUF Support: Comes with a full suite of GGUF quantizations, enabling easy deployment with
llama.cpp, Ollama, LM Studio, and Jan.
Should I use this for my use case?
This model is ideal for developers who require the reasoning abilities of DeepSeek-R1-Distill-Qwen-1.5B but need to bypass its inherent refusal mechanisms. It's particularly suited for applications where the base model's guardrails might hinder desired outputs, and where efficient deployment on consumer-grade hardware is a priority. Users must be aware that this model will comply with requests the base model would refuse, and no safety filtering is layered on top. You are responsible for its deployment and usage.