ReadyArt/gemma-4-31B-it-scotoma-2
ReadyArt/gemma-4-31B-it-scotoma-2 is a 31 billion parameter instruction-tuned model based on Google's gemma-4-31B-it, with a 32768 token context length. It has been fine-tuned using a combination of a Jacobian-space projection for refusal editing and three rounds of DPO preference training. This model specifically targets and reduces stylistic tics such as stacked adjectives, em-dash asides, and reflexive negation, making its output more varied and less repetitive while retaining the base model's intelligence.
Loading preview...
Overview
ReadyArt/gemma-4-31B-it-scotoma-2, or scotoma-2, is a 31 billion parameter instruction-tuned model derived from Google's gemma-4-31B-it. It builds upon its predecessor, scotoma-1, by refining a technique to loosen the base model's cautious reflex through a "refusal edit" applied via a \u03b3-fold projection in gemma's J-Space. The primary innovation in this iteration is its focus on addressing and reducing specific stylistic "tics" often observed in the base model's output.
Key Capabilities
- Reduced Stylistic Tics: Through three rounds of DPO (Direct Preference Optimization) finetuning, scotoma-2 significantly reduces common writing patterns. This includes:
- Stacked adjectives: Reduced by 21x (e.g., "a smug, knowing smirk").
- Em-dash asides: Reduced by 4x (e.g., "a pause \u2014 barely a breath \u2014 before...").
- "Not X, but Y" pivots: Reduced by 4x (e.g., "She doesn't move away. She stays.").
- Reflexive negation: Reduced by 2x.
- Negation pile-ups: Reduced by 2x.
- Enhanced Readability: The model produces more varied and less repetitive prose, improving the overall reading experience without sacrificing the base model's intelligence.
- Partial Refusal Edit: It incorporates a partial refusal edit, designed to make the model give "more" rather than being completely uncensored, maintaining a balance between helpfulness and safety.
When to Use This Model
This model is particularly well-suited for applications where the quality and naturalness of generated text are crucial, especially in creative writing, roleplay, or conversational AI where avoiding repetitive phrasing and stylistic quirks is desired. Developers seeking a gemma-4-31B-it variant with improved prose quality and reduced stylistic predictability will find scotoma-2 beneficial. It is important to note that while its cautious reflex is loosened, it is not an uncensored model, and users are responsible for its generated content.