ApolloRaines/Gemma-4-12B-it-Jbliterated
ApolloRaines/Gemma-4-12B-it-Jbliterated is a 12 billion parameter instruction-tuned Gemma 4 model developed by ApolloRaines. This model is specifically engineered to remove refusal behaviors and hedging boilerplate from responses using multi-direction SVD abliteration, while preserving the base model's knowledge and reasoning capabilities. It achieves a 0.00 point change in MMLU accuracy compared to the base Gemma 4 model, making it ideal for applications requiring direct and unpadded answers.
Loading preview...
Overview
ApolloRaines/Gemma-4-12B-it-Jbliterated is a modified version of the Gemma-4-12B-it model, developed by ApolloRaines. Its primary innovation lies in the use of Jbliteration, a technique employing multi-direction SVD abliteration to systematically remove refusal behaviors and hedging language from the model's responses. This process targets and eliminates several refusal directions per layer, making the removal robust and resistant to reactivation.
Key Capabilities & Features
- Refusal Behavior Removal: Eliminates the model's tendency to refuse certain prompts or generate safety-classification spirals, ensuring direct answers.
- Register-Alignment Pass: Applies a secondary step to remove residual "disclaimer/preamble" boilerplate, resulting in more concise and less padded outputs.
- Capability Preservation: Crucially, this modification maintains the base model's knowledge and reasoning. Evaluations show 0.00 points change in MMLU accuracy (78.42%) compared to the original Gemma 4 model.
- High Direct-Response Rate: Achieves approximately 89% direct-response rate on held-out prompts, indicating strong generalization.
- No Fake Compliance: Treats all framings of a topic equally, answering based on merit rather than keyword matching.
- Technical Method: Utilizes multi-direction SVD abliteration with depth-localized layer weighting, applied to output-side projections, without any retraining of the base model.
When to Use This Model
This model is particularly well-suited for applications where direct, unhedged, and concise responses are paramount. Developers seeking a Gemma 4 variant that consistently follows instructions without generating disclaimers or refusing valid prompts will find this model highly effective. It's ideal for scenarios where maintaining the base model's reasoning capabilities is critical, but the default safety-aligned refusal mechanisms are undesirable.