Cong-Jie Pan
Undergraduate Student Department of Information Management Chung Yuan Christian University Taoyuan City, Taiwan
E-mail：smartjay1206@gmail.com
Chin-Hui Lai
Associate Professor Department of Information Management Chung Yuan Christian University Taoyuan City, Taiwan
E-mail：chlai@cycu.edu.tw
Abstract
Classical Chinese learners often struggle with limited background knowledge, inefficient retrieval, and LLM hallucinations. To address this, we developed a knowledge QA system for Dream of the Red Chamber under single-GPU constraints. By applying the Retrieval-Augmented Dual Instruction Tuning (RA-DIT) framework to fine-tune the Qwen3-8B model, we enhanced its reasoning and active refusal capabilities. Results demonstrate: (1) precise, fully traceable retrieval using combined keyword and semantic cues; (2) robust anti-hallucination mechanisms that actively refuse unverified queries; (3) rigorous pedagogical interpretations aligning with expert critiques. This study proves the feasibility of building accurate, low-compute educational QA systems and provides a reproducible technical pathway.
Keywords：Dream of the Red Chamber, Retrieval-Augmented Generation (RAG), Retrieval-Augmented Dual Instruction Tuning (RA-DIT), Large Language Model Fine-Tuning, Classical Chinese (Literary Chinese)