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

Overview

ApolloRaines/Gemma-4-12B-it-Jbliterated is a 12 billion parameter instruction-tuned model based on the Gemma-4 architecture. Developed by ApolloRaines, this model focuses on eliminating refusal and hedging behaviors commonly found in large language models. It achieves this through a technique called "Jbliteration," which involves precise geometric decomposition of subspaces to remove unwanted directions in the model's weights.

Key Capabilities

• Refusal Removal: Significantly reduces model refusals and hedging, leading to a direct-response rate of ~89% on held-out prompts.
• Capability Preservation: Maintains the base model's MMLU accuracy (78.42%) precisely, indicating no measurable loss in core reasoning abilities.
• No Fake Compliance: Treats all framings of a topic equally, answering on merit rather than keyword-matching.
• Weight Edit Only: Achieves its specialized behavior through weight editing without retraining, preserving the base model's bfloat16 dtype.
• DeepswapLLM Compatibility: Designed to run efficiently even on GPUs with insufficient memory by streaming layers across GPU, RAM, and disk, offering up to 4x faster performance than AirLLM.

Good For

• Applications requiring direct and unbiased answers.
• Use cases where models tend to exhibit excessive safety-classification or refusal behaviors.
• Developers looking for a Gemma-4 variant that prioritizes directness and helpfulness without sacrificing core capabilities.
• Environments with limited GPU memory, leveraging DeepswapLLM for efficient execution.