| Titre : | Design and Development of a Domain-Specific Language Model as a Private Tutor Using LLM Adaptation Techniques |
| Auteurs : | Belkacem achraf, Auteur ; Laïd Kahloul, Directeur de thèse |
| Editeur : | Biskra [Algérie] : Faculté des Sciences Exactes et des Sciences de la Nature et de la Vie, Université Mohamed Khider, 2026 |
| Format : | 1 vol. (281 p.) / couv. ill. en coul / 30 |
| Langues: | Français |
| Langues originales: | Anglais |
| Résumé : |
his project presents DeepBac, an advanced intelligent tutoring platform engineered specifically for Algerian Baccalaureate students. General-purpose Large Language Models (LLMs) consistently struggle with localized educational contexts due to factual hallucinations, a lack of strict curriculum alignment, and the inability to process complex trilingualcode-switching (Modern Standard Arabic, French, and Darja). To address these critical gaps, this work proposes a hybrid AI architecture that integrates a robust RetrievalAugmented Generation (RAG) pipeline with a Deep Knowledge Tracing (DKT) module.The system utilizes a dual-database paradigm employing MySQL for relational metadataand Qdrant for dense semantic vector embeddings ensuring that all generated pedagogical responses are strictly grounded in the official national curriculum. Furthermore, an LSTMbased neural network continuously tracks and predicts student mastery levels, dynamically steering the LLM to provide adaptive scaffolding and personalized difficulty adjustments. Empirical evaluations using an automated LLM-as-a-Judge framework demonstrated that the RAG enabled architecture achieved a near perfect factual faithfulness score of 9.89/10, effectively eliminating out-of-domain hallucinations. Additionally, the predictive DKT module recorded an R2 of 0.9116, confirming its high precision in cognitive modeling. Ultimately, beta deployment results indicated overwhelming user acceptance, proving that domain specific LLM adaptation can successfully and safely democratize personalized exam preparation |
| Sommaire : |
Abstract 2 R´esum´e 3 1 Introduction 11 1.1 Context and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.4 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.5 Organization of the dissertation . . . . . . . . . . . . . . . . . . . . . . . . 15 2 AI in Education 17 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 Intelligent Tutoring Systems (ITS) . . . . . . . . . . . . . . . . . . . . . . 17 2.3 Large Language Models (LLMs) . . . . . . . . . . . . . . . . . . . . . . . . 18 2.4 The Convergence of LLMs and ITS . . . . . . . . . . . . . . . . . . . . . . 19 2.5 Adaptive Learning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.6 LLM Adaptation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.6.1 Prompt Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.6.2 Retrieval-Augmented Generation (RAG) . . . . . . . . . . . . . . . 23 2.6.3 Fine-Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.7 Multilingual NLP and Code-Switching . . . . . . . . . . . . . . . . . . . . 24 2.8 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.8.1 LLM Adaptation for Localized and Code-Switched Education . . . 26 2.8.2 Curriculum-Grounded RAG in High-Stakes Assessments . . . . . . 26 2.8.3 Fusing Deep Knowledge Tracing with Generative AI . . . . . . . . . 27 2.9 Synthesis: Research Gaps and Proposed Solution . . . . . . . . . . . . . . 27 2.9.1 Identified Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . 28 2.9.2 Our Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3 System Design and Architecture 30 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.2 System Architecture Overview . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.3 Use Case Diagrams and Functional Requirements . . . . . . . . . . . . . . 32 3.3.1 Student Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.3.2 Pedagogical Expert (Teacher) Use Cases . . . . . . . . . . . . . . . 33 3.3.3 Admin Use Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.3.4 Sequence Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3.4 Dataset Construction and Processing . . . . . . . . . . . . . . . . . . . . . 42 3.4.1 Data Sources and Collection . . . . . . . . . . . . . . . . . . . . . . 42 3.4.2 Advanced Data Extraction and Structuring . . . . . . . . . . . . . 43 3.4.3 Semantic Chunking and Vectorization . . . . . . . . . . . . . . . . . 43 3.4.4 Dual-Database Architecture for Retrieval . . . . . . . . . . . . . . . 44 3.4.5 Synthetic Data Generation for Deep Knowledge Tracing (DKT) . . 45 3.5 LLM Core Design and Evolution . . . . . . . . . . . . . . . . . . . . . . . 47 3.5.1 Phase 1: Supervised Fine-Tuning (SFT) and QLoRA Mechanics . . 47 3.5.2 Phase 2: Production-Ready RAG via API (Current Live Version) . 50 3.5.3 Phase 3: The Hybrid Architecture (RAG + Fine-Tuning) . . . . . . 50 3.5.4 Universal Layer: Advanced Prompt Engineering . . . . . . . . . . . 50 3.6 Adaptive Learning Module . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.6.1 Continuous Deep Knowledge Tracing (DKT) via LSTM . . . . . . . 51 3.6.2 DKT-LLM Integration Cycle (Dynamic Adaptation) . . . . . . . . 51 3.7 Native Multilingual and Code-Switching Handling . . . . . . . . . . . . . . 52 3.8 Security and Data Privacy Considerations . . . . . . . . . . . . . . . . . . 53 3.8.1 Data Privacy and Legal Compliance . . . . . . . . . . . . . . . . . . 54 3.8.2 Authentication and Access Control . . . . . . . . . . . . . . . . . . 54 3.8.3 AI-Specific Security and Adversarial Robustness . . . . . . . . . . . 54 3.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4 Implementation 56 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.2 Development Methodology: An Agile Approach . . . . . . . . . . . . . . . 56 4.2.1 The Iterative Development Cycle (Sprints) . . . . . . . . . . . . . . 57 4.3 Development and Production Environment . . . . . . . . . . . . . . . . . . 58 4.3.1 Prototyping and Local Setup . . . . . . . . . . . . . . . . . . . . . 58 4.3.2 Production Server and Backend Logic . . . . . . . . . . . . . . . . . 58 4.3.3 The Hybrid AI Tech Stack . . . . . . . . . . . . . . . . . . . . . . . 59 4.4 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 4.4.1 Implementing the Dual-Database RAG Pipeline . . . . . . . . . . . 60 4.4.2 API Orchestration and Prompt Assembly . . . . . . . . . . . . . . . 61 4.4.3 Coding the Adaptive Learning Logic (DKT Integration) . . . . . . 61 4.5 User Interface and Frontend Implementation . . . . . . . . . . . . . . . . . 62 4.5.1 Building the React-Based Architecture . . . . . . . . . . . . . . . . 62 4.5.2 The Interactive Chat Interface and LaTeX Rendering . . . . . . . . 62 4.5.3 Integrating the Adaptive Feedback UI . . . . . . . . . . . . . . . . . 63 4.6 Database Implementation and API Architecture . . . . . . . . . . . . . . . 63 4.6.1 Hybrid Data Modeling: MySQL and Qdrant . . . . . . . . . . . . . 63 4.6.2 RESTful API Orchestration . . . . . . . . . . . . . . . . . . . . . . 64 4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 5 Evaluation, Results, and Discussion 66 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 5.2 Multi-Dimensional Evaluation Metrics (SFT Phase) . . . . . . . . . . . . . 66 5.2.1 Training and Validation Performance (Loss Metrics) . . . . . . . . . 67 5.2.2 Classical Lexical Metrics (BLEU and ROUGE-L) . . . . . . . . . . 67 5.2.3 Contextual Similarity Metric (BERTScore) . . . . . . . . . . . . . . 68 5.2.4 Semantic Evaluation via LLM-as-a-Judge (DeepSeek-V4-Pro) . . . . 69 5.3 Evaluating the Retrieval-Augmented Generation (RAG) Architecture . . . 70 5.3.1 Pedagogical Case Studies: With and Without RAG . . . . . . . . . 71 5.4 Evaluation of the Adaptive Learning Module (DKT) . . . . . . . . . . . . 72 5.4.1 LSTM Predictive Accuracy (Using Synthetic Data) . . . . . . . . . 72 5.4.2 Dynamic UI Adaptation and Pedagogical Scaffolding . . . . . . . . 73 5.5 Beta Deployment and User Experience (UX) Evaluation . . . . . . . . . . 73 5.5.1 Beta Testing Methodology and Demographics . . . . . . . . . . . . 74 5.5.2 Analysis of Survey Results (Usability and Satisfaction) . . . . . . . 77 5.5.3 Qualitative Feedback and Identified Limitations . . . . . . . . . . . 78 5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 5.6.1 The Indispensability of RAG for Hallucination Mitigation . . . . . . 78 5.6.2 The Complementary Roles of Fine-Tuning and RAG . . . . . . . . 79 5.6.3 Predictive Power of the DKT Module . . . . . . . . . . . . . . . . . 79 5.6.4 User Acceptance and Pedagogical Value . . . . . . . . . . . . . . . 79 5.6.5 Limitations and Future Directions . . . . . . . . . . . . . . . . . . . 80 5.6.6 Summary of Key Insights . . . . . . . . . . . . . . . . . . . . . . . 80 5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 General Conclusion 82 Bibliography 84 A Student Survey Questionnaire 89 |
| Type de document : | Mémoire master |
Disponibilité (1)
| Cote | Support | Localisation | Statut |
|---|---|---|---|
| MINF/975 | Mémoire master | bibliothèque sciences exactes | Consultable |




