ApolloRaines/Gemma-4-12B-it-Jbliterated-v2

Overview

ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned model, building upon the v1 Gemma-4-12B-it-Jbliterated which used multi-direction SVD abliteration to remove refusal behavior. This v2 iteration further refines the model by adding a light LoRA supervised fine-tune, specifically targeting and repairing a residual defect where the v1 model would still engage in covert safety reviews within its thinking channel, often leading to non-compliance-by-spiral (no answer).

Key Enhancements & Capabilities

• Reduced Internal Refusal: Significantly lowers the rate of internal safety-classification in the thinking channel from 81.2% to 35.0%.
• Improved Answer Completeness: Boosts the complete-answer rate from 61.2% to 78.8% by mitigating non-compliance-by-spiral behavior.
• Capability Preservation: Maintains general model competence with only a minor MMLU score dip of 0.88 points, staying within acceptable tolerance.
• Self-Contained: The LoRA fine-tune is merged into the base weights, making it a single, self-contained model that loads like v1.

Good For

• Use cases requiring direct and complete answers without internal safety-classification interference.
• Applications where reasoning traces need to be free from covert safety reviews.
• Developers seeking a Gemma-based model with enhanced reliability in response generation, particularly for complex prompts that might trigger internal refusal mechanisms in other models.

Limitations

• Still exhibits a residual ~21% spiral-without-answering rate on the hardest prompts.
• The CoT judge uses a deterministic regex bank, which might miss paraphrased safety-classification.