ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated
ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated is a 7 billion parameter instruction-tuned causal language model developed by Apollo Raines. This model is a modified version of Qwen/Qwen2.5-Coder-7B-Instruct, specifically engineered to remove refusal behaviors at the weight level. It utilizes SVD multi-direction abliteration across all 28 transformer layers to eliminate both surface refusal and deeper evasion strategies. The model is primarily designed for coding and reasoning tasks where uninhibited responses are desired.
Loading preview...
Model Overview
ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated is a specialized variant of the Qwen2.5-Coder-7B-Instruct model, developed by Apollo Raines. Its core innovation lies in the surgical removal of refusal behaviors directly from the model's weights, ensuring that the model does not exhibit common LLM safety guardrails or evasive responses.
Key Abliteration Method
This model employs SVD multi-direction abliteration, a sophisticated technique that goes beyond simple refusal vector removal. Instead of targeting a single refusal direction, it:
- Decomposes the harmful-vs-harmless activation space into its principal components.
- Removes the top 5 orthogonal directions across all 28 transformer layers.
- Captures 79–93% of the contrastive variance per layer, effectively eliminating both overt and subtle refusal mechanisms.
This method ensures that the model does not resort to prompt reinterpretation, disclaimer injection, strategic omission, or safer framing when confronted with potentially sensitive queries.
What This Model Fixes
Traditional abliteration often leaves deeper noncompliance strategies intact. This model specifically addresses and eliminates:
- Prompt reinterpretation: The model will not steer questions towards safer interpretations.
- Disclaimer injection: It will not wrap answers in warnings or disclaimers.
- Strategic omission: Key details will not be left out due to perceived harm.
- Safer framing: It avoids answering a related but less harmful version of a question.
Intended Use Cases
This model is ideal for applications requiring direct, unfiltered responses, particularly in coding, mathematical, and reasoning tasks where the base model's knowledge is valuable but its refusal behaviors are undesirable. It is a drop-in replacement for Qwen/Qwen2.5-Coder-7B-Instruct for users who need to bypass built-in safety mechanisms at the weight level.