luvGPT/luvai-phi3: Specialized Roleplaying LLM
luvGPT/luvai-phi3 is a 4 billion parameter model, fine-tuned from Microsoft's Phi-3-mini-4k-instruct, specifically optimized for engaging in roleplaying conversations with strong character persona consistency. It demonstrates high adaptability to various personality traits defined in the system prompt, making it suitable for diverse narrative and interactive applications.
Key Capabilities
- Persona Consistency: Maintains character traits throughout extended conversations.
- Conversational Flow: Designed to produce natural and engaging dialogue.
- Adaptability: Can adopt a wide range of character personas based on detailed descriptions.
- Efficient Inference: Tested with FP16, 8-bit, and 4-bit quantization, allowing deployment on GPUs with as little as 2.27 GB VRAM.
Important Usage Notes
This model requires a specific prompt template for optimal performance, differing from standard chat formats. Users must define the character's persona and structure messages precisely as Character's Persona: [description]\nUser: [message]\nAssistant:. Failure to follow this format will result in suboptimal outputs.
Training & Performance
The model was fine-tuned using LoRA/deepspeed on a curated dataset of ~13k high-quality roleplaying examples, filtered by a judge model for persona consistency. Training metrics show strong convergence, with token accuracy improving from ~0.48 to ~0.73 and training loss decreasing from ~2.2 to ~1.05 over 3 epochs.
Good For
- Creative Writing: Generating character dialogue and interactive stories.
- Roleplaying Applications: Developing chatbots with distinct personalities.
- Interactive Fiction: Creating dynamic and persona-driven conversational agents.