ConicCat/Gleam-30B

Gleam is a finetune of Muse Glimmer focused on improving roleplay and writing performance using DPO with an emphasis on prose quality and humanlike characters. The dataset consists of approximately 1/3rds writing and 2/3ds roleplay.

How to Get Started with the Model

I recommend grabbing the Q4_K_M gguf and using koboldcpp.

Training Details

Training was conducted on 1xH200 for 5 hours using humanline dpo w/ LoRA.

Datasets

• ConicCat/Lamp_P_Preference: Human revised vs AI writing to improve prose.
• ConicCat/CharCards_Gemma4: Gemma 4 26B > Gemma 4 E2B using the AI2 delta tuning recipe.