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

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 1, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

ApolloRaines/Gemma-4-12B-it-Jbliterated-v2 is a 12 billion parameter instruction-tuned causal language model, a refined version of the Gemma-4-12B-it-Jbliterated base model. Developed by ApolloRaines, this model addresses and significantly reduces internal safety-classification and non-compliance-by-spiral behaviors in reasoning channels. It is optimized for providing complete answers and maintaining general capabilities, making it suitable for applications requiring reliable and direct responses.

Loading preview...

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.