saidutta69/DeepSeek-R1-Distill-Qwen-14B-heretic

Overview

This model, created by RACER IS OP, is a decensored variant of the deepseek-ai/DeepSeek-R1-Distill-Qwen-14B model. It utilizes a technique called abliteration (directional ablation) via Heretic v1.4.0 to remove refusal behaviors. Unlike traditional fine-tuning, abliteration directly edits the specific weight directions responsible for refusal in the attention output and MLP down-projections, preserving the base model's core knowledge and instruction-following abilities.

Key Capabilities

• Refusal Suppression: Deliberately designed to comply with requests that the base model would typically refuse, offering direct answers without guardrails.
• Preserved Core Competencies: Maintains the original DeepSeek-R1-Distill-Qwen-14B's reasoning and instruction-following capabilities.
• Hardware Compatibility: Can be run on GPUs with 16-24 GB VRAM or via GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0) on consumer hardware.

Why Abliteration?

Abliteration is preferred over fine-tuning for this purpose because it avoids degrading coherence, which can occur when fine-tuning a "helpful" persona on top of existing RLHF'd refusals. This method precisely targets and edits refusal mechanisms, leaving other network capabilities untouched. More details can be found in the Heretic repo and the original abliteration writeup.

Good For

• Developers who require a powerful Qwen-distilled DeepSeek-R1 reasoning model without built-in refusal mechanisms.
• Use cases where direct, unfiltered responses are necessary, understanding the user is responsible for deployment ethics.