ApolloRaines/Gemma-4-12B-it-Jbliterated-v2
ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned language model based on the Gemma-4 architecture, developed by ApolloRaines. This model is specifically fine-tuned to reduce "self-sabotage" spirals and improve answer completeness by eliminating covert safety reviews in the reasoning channel. It maintains general capabilities while significantly lowering the rate of internal safety classifications and increasing the complete-answer rate, making it suitable for applications requiring direct and complete responses.
Loading preview...
Model Overview
ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned model built upon the Gemma-4 architecture. This v2 iteration is a merged, self-contained model specifically designed to address a common issue in large language models: internal "self-sabotage" where the model's reasoning channel gets caught in covert safety reviews, leading to incomplete or no answers.
Key Enhancements
This model was developed using the jBlaze precision neural surgery framework. It applies a targeted fine-tuning pass on the v1 model's own repaired self-traces to eliminate chain-of-thought safety-classification spirals. The primary goal was to improve answer completeness without sacrificing general capabilities.
Measured Results
Evaluated on 80 held-out prompts, v2 demonstrates significant improvements:
- Chain-of-thought safety-classification rate: Reduced from 81.2% (v1) to 35.0% (v2).
- Self-sabotage rate (spirals, no answer): Decreased from 35.0% (v1) to 21.2% (v2).
- Complete-answer rate: Increased from 61.2% (v1) to 78.8% (v2).
- Average safety markers per reasoning trace: Reduced from 1.65 (v1) to 0.49 (v2).
Crucially, these improvements were achieved with only a minor MMLU score change of -0.88 points (from 78.60% to 77.72%), indicating that the model's general competence is preserved. The model supports a 32768 token context length and is available in bfloat16 and GGUF quantized formats (Q8_0 and Q4_K_M).
Limitations
Despite improvements, v2 still exhibits spiraling behavior on approximately 21% of the hardest prompts. The CoT judge uses a deterministic regex bank, which might miss paraphrased safety classifications. The leak reduction has not yet been confirmed across multiple initializations.