ArliAI/Gemma-2-9B-ArliAI-RPMax-v1.1 Overview
Gemma-2-9B-ArliAI-RPMax-v1.1 is a 9 billion parameter model from ArliAI's RPMax series, built upon the Gemma-2-9B-IT architecture. This model is uniquely designed for creative writing and roleplay applications, focusing on generating highly varied and non-repetitive content. Its training methodology specifically curates diverse datasets, ensuring that no two entries repeat characters or situations, which helps the model avoid 'in-bred' or predictable responses.
Key Capabilities
- Highly Creative Generation: Engineered to produce diverse and imaginative narratives.
- Non-Repetitive Output: Training focuses on minimizing repetition sickness, a common issue in other roleplay models.
- Adaptable Character & Situation Handling: Capable of understanding and appropriately responding to a wide range of characters and scenarios without latching onto specific personalities.
- Optimized for Roleplay: Specifically fine-tuned for engaging and dynamic roleplaying experiences.
Training Details
The model underwent a 1-epoch training duration of approximately 2 days on 2x3090Ti, utilizing QLORA with 64-rank 128-alpha, resulting in ~2% trainable weights. It was trained with a sequence length of 4096 and a learning rate of 0.00001, with a gradient accumulation of 32.
Good For
- Creative Writing: Generating unique stories, dialogues, and descriptive text.
- Roleplaying Scenarios: Engaging in dynamic and varied character interactions.
- Applications Requiring Novelty: Use cases where avoiding repetitive language and ideas is crucial.
Prompt Format
This model uses the Gemma Instruct Prompt Format. Character descriptions should be included in the first user turn, enclosed by <instructions> and <end_of_instructions> tags, similar to Mistral models.