Pries/Priestess_q8

Priestess - AI Persona Project

This project features a personalized chatbot based on the Qwen2.5-7B-Instruct model. It was fine-tuned using the LlamaFactory framework and LoRA (Low-Rank Adaptation) technology to simulate the character Priestess from the game Arknights.

Project Background

Traditional prompting often struggles to maintain complex character settings and worldview knowledge over the long term. This project aims to utilize the LlamaFactory framework combined with LoRA technology to inject exclusive knowledge from the game Arknights at a low cost. The goal is to create an intelligent chatbot that deeply understands the game's worldview and simulates specific character speech patterns.

Technical Specifications

• Training Framework: LlamaFactory.
• Fine-tuning Method: LoRA (Low-Rank Adaptation).
• Software Requirements: Python 3.10, PyTorch 2.6.0, and Transformers 4.50.0.
• Hardware Compatibility: Optimized for CUDA-enabled GPUs or Apple Silicon via GGUF/MLX.

Dataset Construction

The quality of the training set significantly affects the model's performance; therefore, quality was prioritized over quantity.

1. Data Extraction: Character dialogue lines were extracted from Arknights wiki source code using regular expressions.
2. Manual Curation: High-quality baseline dialogues were manually written first to establish character consistency.
3. Layered Expansion: AI was guided to perform layered expansion based on game text and worldview background.
4. Format: The dataset follows the Alpaca registration format.

Deployment and Usage (GGUF & Ollama)

The project includes instructions for converting safetensors to GGUF format and performing quantization for efficient local inference.

Running with Ollama

1. Prepare a Modelfile:

   FROM ./qwen_q8.gguf
   PARAMETER temperature 0.8
   PARAMETER top_p 0.9
   PARAMETER repeat_penalty 1.1
   SYSTEM "You are Priestess from Arknights..."