MyGemmaNPC

Model Card for MyGemmaNPC

Model Overview

MyGemmaNPC is a lightweight, fine-tuned version of Gemma-3-270M-IT, specifically optimized for game character dialogue (NPCs). It utilizes a 32K token context window and is designed for low-latency edge deployment.

Architectural Specifications

• Base Model: google/gemma-3-270m-it
• Training Method: Supervised Fine-Tuning (SFT) via TRL
• Precision: bfloat16 (default) / q4_0 (quantized for 240MB VRAM)

Intended Use

• Primary Use: Roleplay dialogue generation for non-player characters in interactive storytelling.
• Out-of-Scope: This model is not designed for fact-checking, code generation, or medical advice.

Training Details

• Hardware: Trained on a single-node GPU cluster.
• Optimizer: AdamW with a learning rate of 2e-5.
• Data: Fine-tuned on a custom dataset of character dialogues.

Bias & Limitations

As a 270M parameter model, it may exhibit "common knowledge" gaps and a higher propensity for hallucination on niche technical topics compared to the larger Gemma 3 variants.