ConicCat/Gleam-30B
ConicCat/Gleam-30B is a 30 billion parameter language model developed by ConicCat, fine-tuned from Muse Glimmer. It utilizes DPO with a 131072 token context length to enhance roleplay and writing performance. This model emphasizes prose quality and humanlike character interactions, making it suitable for creative writing and interactive narrative generation.
Loading preview...
Gleam-30B Overview
Gleam-30B is a 30 billion parameter language model developed by ConicCat, building upon the Muse Glimmer architecture. It has been specifically fine-tuned using DPO (Direct Preference Optimization) to significantly improve its capabilities in roleplay and creative writing tasks. A key focus during its development was enhancing prose quality and generating more humanlike character interactions.
Key Capabilities
- Enhanced Roleplay: Optimized for generating dynamic and engaging character interactions.
- Improved Writing Quality: Focuses on producing high-quality prose, making it suitable for various creative writing applications.
- Humanlike Characters: Designed to create more believable and nuanced character responses.
- Extensive Context Window: Features a substantial 131072 token context length, allowing for longer and more coherent narratives.
Training Details
The model was trained for 5 hours on 1xH200 using humanline DPO with LoRA. The training dataset was composed of approximately one-third writing data and two-thirds roleplay data, leveraging specific datasets like ConicCat/Lamp_P_Preference for prose improvement and ConicCat/CharCards_Gemma4 for character-focused tuning.
Good For
- Creative Writing: Generating stories, dialogues, and descriptive text.
- Interactive Fiction & Roleplaying Games: Powering character responses and narrative progression.
- Content Generation: Creating engaging and high-quality textual content with a focus on narrative and character depth.