saidutta69/DeepSeek-R1-Distill-Qwen-7B-heretic
The saidutta69/DeepSeek-R1-Distill-Qwen-7B-heretic is a 7.6 billion parameter language model, derived from deepseek-ai/DeepSeek-R1-Distill-Qwen-7B, with a 32768 token context length. This variant has undergone 'abliteration' using Heretic v1.4.0 to suppress refusal behavior through targeted weight edits, rather than fine-tuning. It maintains the base model's strong reasoning and instruction-following capabilities, making it suitable for developers seeking a capable 7B model that provides direct answers without refusal guardrails.
Loading preview...
Model Overview
saidutta69/DeepSeek-R1-Distill-Qwen-7B-heretic is a 7.6 billion parameter language model based on the deepseek-ai/DeepSeek-R1-Distill-Qwen-7B architecture, featuring a 32768 token context length. Its primary distinction lies in the removal of refusal behaviors, achieved through a process called abliteration using Heretic v1.4.0.
Key Capabilities & Differentiators
- Decensored Output: Refusal behavior is suppressed via targeted weight edits to attention output and MLP down-projections, allowing the model to answer directly without guardrails.
- Preserved Core Abilities: Unlike traditional fine-tuning for decensoring, abliteration leaves the base model's knowledge and instruction-following capabilities largely intact, ensuring coherence and reasoning quality.
- Optimized for Reasoning: Inherits the strong reasoning capabilities of the DeepSeek-R1-Distill-Qwen-7B base model, making it suitable for chain-of-thought applications.
- Hardware Friendly: Designed to run efficiently on consumer-grade GPUs, with various GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0) available to fit different VRAM capacities (e.g., 8-12 GB GPUs).
Intended Use Cases
This model is ideal for developers who require a capable 7B reasoning model that provides direct answers without inherent refusal mechanisms. It's particularly suited for scenarios where the base model's refusal guardrails are undesirable, and users are responsible for managing the output. It can be deployed using llama.cpp, transformers, Ollama, LM Studio, Jan, vLLM, and SGLang.