chimbiwide/Gemma4NPC-E4B
Gemma4NPC-E4B is a 7.9 billion parameter language model developed by chimbiwide, based on the Gemma4-E4B architecture. Fine-tuned with a 32768-token context length, this model is specifically optimized for game-focused NPC roleplaying. It excels at generating conversational and authentic responses for non-player characters, referencing character backgrounds and current surroundings naturally.
Loading preview...
Model Overview
chimbiwide/Gemma4NPC-E4B is a 7.9 billion parameter model built upon the new Gemma4-E4B base, specifically fine-tuned for game-focused Non-Player Character (NPC) roleplaying. It leverages the RolePlay-NPC-Quest dataset to enhance its ability to generate dynamic and context-aware NPC interactions.
Key Capabilities
- Game-focused NPC Roleplaying: Designed to create realistic and engaging NPC dialogues.
- Contextual Awareness: Generates responses that naturally reference character backgrounds, bios, and current locations.
- Instruction Following: Shows significant improvement in instruction-following compared to previous Gemma3/3n models, allowing for detailed roleplaying instructions.
- Conversational Flow: Produces authentic and conversational responses, reacting to player input.
Training Details
The model was trained as an r=32, alpha=64 LoRA adapter over 2 epochs on the RolePlay-NPC-Quest dataset. Training utilized an 80GB vRAM A100, with a learning rate of 1e-4, an effective batch size of 24, and a cosine learning rate scheduler with a 500-step warmup. Gradient clipping was set at 1.0. Despite observing higher loss and grad_norm during text-only training, as noted by Unsloth for Gemma4, the model demonstrates strong performance.
Recommended Usage
For optimal performance, inference should use temperature = 1.0, top_p = 0.95, and top_k = 64. The model is designed to be prompted with specific system prompts for roleplay mode, including character details, location, and roleplaying instructions, optionally with a quest objective. It is recommended to start interactions with a user prompt of "Greetings" to initiate an NPC-like greeting.