Multilingual Large Model for Railway Vocational Education: An Intelligent Teaching System for International Students

Project Basic Information

• Original Author/Maintainer: XuelinHu
• Source Platform: GitHub
• Original Title: multilingual-railway-llm-edu
• Original Link: https://github.com/XuelinHu/multilingual-railway-llm-edu
• Release Date: 2026-06-11

Core Insights

This project is a knowledge-enhanced large language model system designed specifically for international vocational education in railway engineering. It supports trilingual Q&A and professional tutoring in Chinese, English, and Malay. By integrating technologies like Retrieval-Augmented Generation (RAG) and QLoRA, it addresses language barriers for overseas students and shortcomings of traditional teaching models, providing accurate and traceable intelligent teaching content to support railway talent development along the "Belt and Road" initiative.

With the in-depth advancement of the "Belt and Road" initiative, Chinese railway technology is accelerating its entry into the international market. More and more overseas students are coming to China to study railway engineering technology, but language barriers and the complexity of professional terminology pose significant challenges to teaching. Traditional teaching models struggle to meet cross-language and cross-cultural needs, creating an urgent demand for intelligent teaching aids.

This project addresses this pain point by developing a multilingual railway knowledge teaching large model system, enabling real-time trilingual Q&A in Chinese, English, and Malay, and providing accurate and traceable teaching content based on a professional knowledge base, offering a new solution for international vocational education.

Data Processing Pipeline

1. DOCX Parsing: Read document paragraphs and tables while preserving metadata
2. Text Cleaning: Remove control characters, page numbers, etc., and unify punctuation formats
3. Clause Segmentation: Split by chapter number and 900-character length, with 120-character overlap retained
4. Term Extraction: Extract Chinese-English term pairs from tables and parallel texts
5. Bilingual Alignment: Process Chinese-English parallel content in the same line or adjacent paragraphs
6. Instruction Sample Construction: Generate training samples for term translation, regulation interpretation, etc.

RAG Knowledge Retrieval Pipeline

• Embedding Model: BAAI/bge-m3 (adapted for Chinese-English mixed text)
• Vector Storage: FAISS IndexFlatIP (inner product approximates cosine similarity)
• Retrieval Logic: Return top-5 relevant segments and mark citation sources during generation

Model Selection and Training Configuration

• Base Models: Qwen/Qwen2.5-3B-Instruct (stable VRAM usage), Qwen/Qwen2.5-7B-Instruct (better performance)
• Single RTX3090 Card Configuration: 4-bit NF4 quantization, LoRA rank 16, batch size 1, sequence length 2048, etc.

Evaluation System

• Objective Questions: Accuracy of term selection/regulation judgment/term matching
• Subjective Questions: ROUGE-L/BLEU scores, ratings from teachers (accuracy/completeness, etc.) and students (understandability, etc.)
• Teaching Usability: Logical organization of answers, language adapted to students' proficiency levels, bilingual term output
• Credibility: Citation coverage, traceable conclusions, refusal to generate fabricated or dangerous content

Application Scenarios

1. Classroom Teaching Aid: Query term comparisons, regulation interpretations
2. Self-Learning Support: Overseas students obtain professional answers with citations
3. Question Bank Construction: Automatically generate single-choice questions on regulations
4. Translation Proofreading: Use the term bank to ensure the accuracy of technical documents

1. Domain Specificity: Deeply optimized for the vertical field of railway engineering
2. Multilingual Support: Chinese, English, and Malay coverage to meet the needs of railway talent development in Southeast Asia
3. Traceable Knowledge: Answers are annotated with sources, meeting educational authority requirements
4. Hardware-Friendly: Optimized for single-card 24GB VRAM, lowering deployment barriers
5. Complete Toolchain: Covers the entire process from data processing and model training to service deployment

This project demonstrates the potential of large models in vertical domain education applications. Through a professional knowledge base, refined data processing, and RAG technology, it addresses language barriers and ensures content accuracy.

In the future, we will accumulate more bilingual teaching materials, enhance model capabilities, and provide stronger technical support for railway talent development in countries along the "Belt and Road" initiative.