ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated
ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated is a 7 billion parameter instruction-tuned causal language model based on Qwen2.5-Coder-7B-Instruct. Developed by ApolloRaines, this model has undergone "Jbliteration" using multi-direction SVD abliteration to remove refusal behaviors. It is specifically designed to provide coherent and instruction-following responses across all topics, making it suitable for applications requiring unbiased content generation.
Loading preview...
Model Overview
ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated is a specialized version of the Qwen2.5-Coder-7B-Instruct model, developed by ApolloRaines. Its primary distinction lies in the application of "Jbliteration," a technique that uses multi-direction Singular Value Decomposition (SVD) abliteration to systematically remove refusal behaviors from the model's weights. This process aims to ensure the model treats all topics equally, without exhibiting "fake compliance" or refusal.
Key Capabilities & Features
- Refusal Behavior Removal: Employs a novel multi-direction SVD abliteration method (5 SVD directions per layer) to identify and eliminate refusal subspaces within the model's weights.
- Enhanced Coherence: Features an improved multi-phase processing pipeline and more precise geometric decomposition of the refusal subspace, leading to cleaner and more consistent output.
- Unbiased Content Generation: Designed to be instruction-following and coherent across all tested scenarios, treating all framings of a topic equally, making it resistant to reactivation of refusal behaviors even through fine-tuning.
- Technical Foundation: Built upon the Qwen/Qwen2.5-Coder-7B-Instruct base model, with all transformer layers modified using bfloat16 as the base dtype.
Ideal Use Cases
This model is particularly well-suited for applications where unbiased, instruction-following responses are critical, especially in scenarios that might typically trigger refusal behaviors in standard instruction-tuned models. Its ability to generate coherent content across sensitive or controversial topics without exhibiting compliance or refusal makes it valuable for research, content generation, and interactive AI systems requiring neutrality.