CrucibleLab/Gemma-4-31B-Loki-Scotoma-V2.0-Swa-4096
CrucibleLab's Gemma-4-31B-Loki-Scotoma-V2.0-Swa-4096 is a 31 billion parameter language model based on ReadyArt/gemma-4-31B-it-scotoma-2, fine-tuned on the Loki V2 dataset. It features a two-stage sliding window attention (SWA) expansion to 4096 tokens, improving long-context retrieval by 10-15% and enhancing local continuity in multi-turn roleplay. This model is specifically designed for deeply immersive roleplay and narrative experiences, prioritizing consistent long-form storytelling, narrative authority, and creative prose over generic chatbot behavior.
Loading preview...
Overview of Gemma-4-31B-Loki-Scotoma-V2.0-Swa-4096
This model, developed by CrucibleLab, is a 31 billion parameter language model built upon ReadyArt/gemma-4-31B-it-scotoma-2, which itself is a fine-tune of google/gemma-4-31B-it. It has been further trained on the Loki V2 dataset to specialize in immersive roleplay and narrative generation. A key differentiator is its two-stage Sliding Window Attention (SWA) expansion, progressively extended from 1024 to 4096 tokens.
Key Capabilities
- Enhanced Long-Context Retrieval: Demonstrates an approximate 10%-15% improvement in BABILong Needle Recall over standard SWA layers, making it more effective for long-context scenarios.
- Superior Narrative Continuity: Significantly sharper local continuity and cross-paragraph reasoning, particularly beneficial for multi-turn roleplay.
- Immersive Roleplay: Designed to act as a Dungeonmaster, Story Author, or Creative Narrator, focusing on tone, voice, continuity, and world fidelity rather than generic chatbot responses.
- Narrative Authority: Aims for consistent long-form storytelling with clear narrative roles, preventing role confusion and character breaks.
- Thematic Adaptability: Trained on diverse fictional universes (Anime, Fantasy, Sci-Fi) to capture specific styles, tones, and narrative aesthetics.
- Flexible Narrative Modes: Supports distinct modes like 'Dungeonmaster', 'Storywriter', and 'Storyteller' to guide its narrative approach, along with tone tags ([Tone: Neutral], [Tone: Dark], [Tone: Mature]) and specific anatomy tags for diverse creative writing.
Good For
- Roleplay enthusiasts seeking deeply immersive and consistent experiences.
- Story-driven prompt engineers who require models to act as creative partners.
- Writers utilizing LLMs as advanced narrative tools for long-form content.
- Users who prioritize rich, detailed stories and narrative control over general assistance.