saidutta69/Qwen2.5-14B-Instruct-heretic

Qwen2.5-14B-Instruct-heretic: Decensored Qwen2.5-14B-Instruct

This model is a 14.8 billion parameter variant of Qwen/Qwen2.5-14B-Instruct, created by RACER IS OP using the Heretic v1.4.0 abliteration method. Unlike traditional fine-tuning, abliteration directly edits specific weight directions responsible for refusal, leaving the base model's core capabilities and instruction-following largely intact. This approach aims to provide a model that answers directly without refusing prompts that the base model would typically decline.

Key Characteristics & Performance

• Refusal Suppression: Achieves a significant reduction in refusals, dropping from 98/100 to 14/100 on adversarial prompts, while maintaining a low KL divergence of 0.0600 from the base model's output distribution.
• Capability Preservation: The targeted weight edits ensure that the original knowledge and instruction-following abilities of Qwen2.5-14B-Instruct are preserved.
• Hardware Compatibility: Optimized with GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0) to run efficiently on consumer GPUs, including those with 8GB to 24GB VRAM, and CPU-only setups.

Use Cases

This model is ideal for developers who require a powerful instruction-following model that provides direct answers without built-in refusal mechanisms. It is particularly suited for applications where the base model's refusal guardrails are undesirable, offering strong reasoning and instruction-following for a wide range of tasks. Users are responsible for its deployment, as it lacks safety filtering.