saidutta69/Qwen2.5-14B-Instruct-heretic
The saidutta69/Qwen2.5-14B-Instruct-heretic is a 14.8 billion parameter instruction-tuned causal language model, a decensored variant of Qwen/Qwen2.5-14B-Instruct. Developed by RACER IS OP using the Heretic abliteration method, this model suppresses refusal behavior via targeted weight edits rather than fine-tuning. It retains the base model's strong reasoning and instruction-following capabilities, making it suitable for developers requiring direct answers without refusal guardrails.
Loading preview...
Model Overview
The saidutta69/Qwen2.5-14B-Instruct-heretic is a 14.8 billion parameter instruction-tuned model, derived from Qwen/Qwen2.5-14B-Instruct. Its primary distinction is the removal of refusal behaviors through a process called abliteration (directional ablation) using Heretic v1.4.0, rather than traditional fine-tuning. This method involves targeted weight edits to the attention output and MLP down-projections, preserving the base model's core knowledge and instruction-following abilities.
Key Characteristics
- Decensored Behavior: Significantly reduced refusal rates (14/100 adversarial prompts compared to 98/100 for the base model).
- Preserved Capabilities: Maintains the strong reasoning and instruction-following of the original Qwen2.5-14B-Instruct.
- Targeted Modification: Abliteration ensures minimal impact on the model's overall coherence and capabilities, with a low KL divergence of 0.06 from the base model's output distribution.
- Hardware Friendly: Can be run on GPUs with 16-24 GB VRAM or via GGUF quantizations (Q4_K_M/Q5_K_M) on consumer hardware.
Ideal Use Cases
- Developers seeking a large Qwen2.5 instruct model that provides direct answers without refusal guardrails.
- Applications requiring consistent, non-refusal responses to a broad range of prompts.
- Environments where the base model's knowledge and instruction-following are valued, but its inherent safety filters are undesirable.
Note: This model is intended for responsible use. It will comply with requests the base model would refuse, including potentially harmful ones, and lacks additional safety filtering. Users are responsible for its deployment and moderation.