sinonchum/cyberdiviner-gemma-3-1b

CyberDiviner Gemma 3 1B

CyberDiviner Gemma 3 1B is a Chinese format-alignment fine-tune of Gemma 3 1B for the CyberDiviner app.

The model is optimized for short, structured, on-device divination-style responses across:

• Oracle-style readings
• Tarot interpretations
• Liuyao / Zhouyi readings
• Face-reading text generation

Intended Use

This model is intended as an application-specific text generator for CyberDiviner. It focuses on stable output format, concise Chinese prose, and reduced prompt or format leakage.

It is not intended for medical, legal, financial, or other high-stakes decision making.

Base Model

• Base model: google/gemma-3-1b-it
• Fine-tuning method: QLoRA SFT
• Output format: merged Hugging Face model

Deployment Notes

This repository contains the merged Hugging Face model. The Android app does not load this directory directly.

For CyberDiviner on-device inference, this model still needs to be converted into a LiteRT-LM compatible int4 .task file, then validated on a physical Android device.

Evaluation Focus

Acceptance should prioritize:

• Fixed section formats
• No repeated sentence loops
• No prompt leakage
• No fallback wording such as "local prophet" or implementation details
• No raw feature fields, pixel values, ratios, or numeric code strings in user-facing output
• Stable latency and memory behavior on device

License

This model is derived from Gemma and is subject to the Gemma license terms. Users are responsible for complying with the base model license and applicable distribution requirements.