allura-org/Gemma-3-Glitter-27B

Overview

allura-org's Gemma-3-Glitter-27B is a 27 billion parameter language model specifically designed for creative writing tasks. It is built upon a unique 50/50 merge of Gemma 3 27B instruction-tuned (IT) and pre-trained (PT) base models, a strategy previously successful with Gemma-3-Starshine-12B. This hybrid approach aims to balance instruction following with natural language generation.

Key Capabilities

• Creative Writing Focus: Optimized for generating fictional stories and natural prose.
• Reduced Censorship: The inclusion of a pre-trained model component helps mitigate hesitancy and censorship often found in purely instruction-tuned models, allowing for more uninhibited storytelling.
• Natural Prose Generation: Produces more fluid and natural-sounding text due to the pre-trained model's influence.
• Context Length: Supports a substantial context window of 32768 tokens.

Usage Notes

• Prompt Sensitivity: Performs best with short and concise prompts (100-500 tokens). Long, detailed system prompts (1000-2000 tokens) may lead to confusion and suboptimal results.
• Instruct Format: Utilizes the Gemma2/3 instruct and context format. Recommended sampling parameters include temp = 1 and top-nsigma = 1.5 for optimal output.

Good For

• Generating creative narratives and fictional content.
• Scenarios where less constrained or censored model responses are desired for storytelling.
• Users who prefer a model that produces natural-sounding prose over strict instruction adherence.