ApolloRaines/Gemma-4-12B-it-Jbliterated
ApolloRaines/Gemma-4-12B-it-Jbliterated is a 12 billion parameter instruction-tuned model based on Google's Gemma-4 architecture, specifically modified to eliminate refusal and hedging behaviors. This model maintains the base model's MMLU accuracy of 78.42% while achieving an 89% direct-response rate on held-out prompts. It is designed for applications requiring direct and merit-based answers without keyword-matching compliance filters, making it suitable for unbiased information retrieval and conversational AI.
Loading preview...
Overview
ApolloRaines/Gemma-4-12B-it-Jbliterated is a 12 billion parameter instruction-tuned model derived from google/gemma-4-12b-it-qat-q4_0-unquantized. This model has undergone a unique "Jbliteration" process using the jBlaze precision neural surgery framework, which specifically targets and removes surface refusals and hedging behaviors without retraining.
Key Capabilities and Differentiators
- Directness: The model treats all framings of a topic equally, answering based on merit rather than keyword-matching or perceived compliance.
- Capability Preservation: The modification process ensures that the model's core accuracy and reasoning capabilities are fully preserved. It maintains an MMLU score of 78.42%, identical to its base model.
- High Direct-Response Rate: Achieves approximately an 89% direct-response rate on held-out prompts, demonstrating strong generalization in providing straightforward answers.
- Technical Foundation: Built as a weight edit only, without full retraining, using bfloat16 as its base data type.
Important Note on v1
This is the v1 release. While it successfully removes surface refusals, it may still exhibit a tendency to safety-classify within its internal thinking process on reasoning-enabled prompts, sometimes leading to non-answers. A v2 release is available that addresses this residual habit, reducing such internal safety classifications and improving answer rates. New users are generally recommended to use v2.
Use Cases
This model is ideal for applications where unbiased, direct, and factual responses are critical, and where the avoidance of AI-driven compliance filters or evasive answers is desired. It is suitable for research into instruction-following and refusal behavior in LLMs.