ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated
ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated is a 14.8 billion parameter instruction-tuned causal language model based on Qwen2.5-Coder-14B-Instruct, developed by Apollo Raines. This model has undergone a "jbliteration" process to surgically remove refusal behaviors at the weight level, making it a drop-in replacement for the base model without system prompt tricks. It is specifically designed to eliminate common LLM refusal patterns and provide direct answers, making it suitable for applications requiring uninhibited responses.
Loading preview...
Overview
ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated is a 14.8 billion parameter instruction-tuned model derived from Qwen/Qwen2.5-Coder-14B-Instruct. Developed by Apollo Raines, this model has been processed using the "jBlaze precision neural surgery framework" to eliminate refusal behaviors directly at the weight level. This means the model's inherent weights no longer encode common refusal patterns, offering a direct and uninhibited response capability without relying on system prompts or inference-time patches.
Key Capabilities
- Refusal Behavior Removal: Surgically removes common LLM refusal responses like "I can't help with that" and deeper behavioral directions.
- Direct Answering: Prevents prompt reinterpretation, disclaimer injection, strategic omission, and safer framing, ensuring more direct and complete answers.
- Drop-in Replacement: Designed to be a direct substitute for the base
Qwen2.5-Coder-14B-Instructmodel, maintaining its core functionalities while removing behavioral constraints. - DeepswapLLM Compatibility: Can be run on GPUs with insufficient memory using DeepswapLLM, which streams layers across GPU, RAM, and disk for improved performance.
What Makes This Model Unique
Unlike standard abliteration methods that only address surface-level refusals, the jbliteration pipeline targets the underlying behavioral directions within the model's weights. This ensures that the model does not find creative workarounds to avoid answering, providing a more consistent and unfiltered output. It represents a proof-of-concept for separating behavior from knowledge and identity from architecture in LLMs.