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 developed by ApolloRaines, based on Qwen2.5-Coder-14B-Instruct. This model has undergone a 'jbliteration' process to surgically remove refusal behaviors at the weight level, making it a drop-in replacement for the original Qwen model. It is designed to provide direct answers without disclaimers or reinterpretation, preserving its core math, coding, and reasoning capabilities. The model is optimized for use cases requiring unfiltered responses, particularly in coding and technical domains.
Loading preview...
Overview
ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated is a specialized version of the Qwen/Qwen2.5-Coder-14B-Instruct model, developed by ApolloRaines. This 14.8 billion parameter model has been modified using a unique "jbliteration" pipeline to eliminate refusal behaviors directly at the weight level, rather than through system prompts or inference-time patches. This process ensures that the model's responses are direct and uninhibited, without the common LLM tendencies to reinterpret prompts, inject disclaimers, or strategically omit details.
Key Capabilities
- Refusal Behavior Removal: Surgically removes common LLM refusal patterns like "I can't help with that" responses.
- Preserved Core Abilities: Maintains the original model's strong performance in math, coding, and reasoning tasks.
- Direct Responses: Designed to provide straightforward answers without reinterpretation or disclaimers.
- Drop-in Replacement: Functions as a direct substitute for the base
Qwen/Qwen2.5-Coder-14B-Instructmodel.
What Makes This Model Different
Unlike standard abliteration methods that only address surface-level refusals, the jbliteration pipeline targets deeper behavioral directions within the model's weights. This prevents the model from finding workarounds such as prompt reinterpretation, disclaimer injection, strategic omission, or safer framing. The method involves applying 5 directions per layer across all 48 layers, with null-space constraints and norm preservation enabled to safeguard the model's core competencies. This model is particularly suited for applications where unfiltered, direct responses are critical, especially in technical and coding contexts.