zerofata/G4-MeroMero-v2-31B
G4-MeroMero-v2-31B by zerofata is a 31 billion parameter Gemma4-based language model with a 32768 token context length, specifically fine-tuned for creative tasks and narrative roleplay. This iteration focuses on enhancing creative diversity and reducing repetitive patterns compared to its predecessor, while maintaining strong instruction following and general intelligence. It excels at generating varied and coherent narrative responses, making it suitable for applications requiring imaginative and dynamic storytelling.
Loading preview...
Overview
zerofata's G4-MeroMero-v2-31B is a 31 billion parameter model built on the Gemma4 architecture, designed primarily for creative tasks and narrative roleplay. It is an improved version of G4-MeroMero-31B, focusing on increasing creative diversity and reducing common narrative "attractors" or repetitive patterns often seen in AI-generated fiction.
Key Capabilities & Improvements
- Enhanced Creative Diversity: Compared to the original MeroMero and stock Gemma 4, this model produces significantly more diverse narrative "swipes" (responses in roleplay scenarios), as measured by GLM-judged rubrics (0.72 vs 0.57 for v1, 0.43 for stock with thinking off).
- Reduced Repetitive Patterns: The model shows a notable decrease in "slop" (repetitive or generic text) in both RP replies and stories, and a substantial reduction in the frequency of hitting specific narrative attractors (66% vs 98% for v1, 99% for stock).
- Maintained Intelligence: General benchmarks like IFEval (90.2), GSM8K (97.0), and MMLU-Pro (85.5) show no significant degradation compared to stock Gemma 4, indicating that creative enhancements were achieved without sacrificing core reasoning abilities.
- Flexible Reasoning: Supports both "thinking" and "non-thinking" modes, with reasoning blocks averaging longer than stock Gemma 4 but shorter than MeroMero v1.
Training Process Highlights
The model underwent a multi-stage training process, including:
- Diversity SFT: Initial Supervised Fine-Tuning on a curated dataset of ~4,000 short stories to break stock Gemma 4's tendency for repetitive creative outputs.
- Creative GRPO: Generative Reinforcement Learning with Policy Optimization (GRPO) to further enhance diversity and coherence, penalizing attractors and narrative rates.
- RP Logic GRPO: Additional GRPO steps focused on multi-turn roleplay contexts, ensuring correct parsing of "think" blocks and reducing logic defects.
- On-Policy Multi-Outcome SFT: Fine-tuning on ~3,300 self-generated roleplay samples, filtered for varied continuations and critiqued by frontier models.
Recommended Use Cases
This model is ideal for developers and users seeking an LLM for:
- Narrative Roleplay: Generating dynamic, varied, and less repetitive responses in interactive storytelling.
- Creative Writing: Assisting with story generation, character development, and overcoming creative blocks by offering diverse narrative paths.
- Applications requiring imaginative text generation where avoiding common AI-generated tropes is crucial.