A peft-optimized language model for curriculum alignment and educational coherence

Abstract

This paper presents the development and fine-tuning of a lightweight natural language processing (NLP) model, optimized through Parameter-Efficient Fine-Tuning (PEFT) strategies, aimed at enhancing curriculum alignment and educational coherence. Using the LoRA (Low-Rank Adaptation) method, the transformer-based model microsoft/phi-2 was adapted to analyze the consistency between curricular objectives, lecture content, laboratory guides, and final examination items. The project addresses the critical challenge of assessing content coverage and thematic coherence, a difficulty extensively discussed in educational research. The primary objectives were twofold: (i) to evaluate the degree of curricular coverage and logical alignment between teaching and testing materials, and (ii) to develop an AI-based assistant capable of generating curriculum-aware, pedagogically relevant responses to support learners. Romanian and Russian course materials were automatically translated into English using the Helsinki-NLP Opus-MT models, ensuring multilingual training adaptability. Data preparation involved structuring datasets in JSONL format, advanced tokenization, and adaptation for causal language modeling. To enable efficient deployment on resource-constrained hardware, the final model was quantized to 8-bit precision utilizing the bitsandbytes optimization library. By leveraging PEFT methods such as LoRA, the fine-tuning process achieved substantial reductions in computational resource usage without degrading output quality.Model evaluation, conducted through a multilingual interactive script with real-time translation, confirmed the model’s ability to diagnose curriculum gaps, redundancies, and inconsistencies while maintaining high coherence and relevance in response generation. The proposed methodology contributes to educational NLP research by offering a scalable, resource-efficient approach to training specialized AI assistants, directly addressing the need for systematic evaluation frameworks in content coherence analysis.