🧠 LiSoViMa: A Lightweight and Specialized Virtual Assistant for STEM MCQA

📍 EPFL – Master in Data Science, Year 1 (2025)
👥 Team: Matthias Wyss, Sofia Taouhid, Vincent Fiszbin, Lina Sadgal
📄 Final Report: Report
🔗 GitHub Repository: GitHub

---

LiSoViMa is a lightweight language model-based assistant tailored to STEM multiple-choice question answering (MCQA). Built on the Qwen3-0.6B-Base model, it evaluates the trade-offs and synergies between various fine-tuning and optimization techniques for educational tasks.

The project implements and evaluates:

• 🧩 LoRA fine-tuning for accurate MCQA on science and math datasets
• 📉 4-bit QLoRA quantization to reduce memory usage while maintaining performance
• 📚 RAG-based enhancement using textbook-based retrieval with a FAISS index and markdown pipeline
• 🧠 Direct Preference Optimization (DPO) for aligning model responses with human preferences

Each variant is benchmarked across six STEM datasets, considering accuracy, memory efficiency, and alignment quality.

---

We also explore:

• Construction and formatting of textbook corpora for retrieval
• Dataset curation from public STEM QA sources (e.g., MMLU-STEM, SciQ, ARC)
• Human and synthetic preference data creation
• LoRA merging, checkpoint management, and reproducibility

---

🛠 Tools & Libraries:

• Python
• Hugging Face Transformers
• PyTorch
• FAISS
• LaTeX (for report and derivations)

📐 Techniques:

• LoRA-based fine-tuning
• Retrieval-Augmented Generation (RAG)
• 4-bit Quantization (QLoRA, GPTQ)
• Direct Preference Optimization (DPO)
• MCQA Evaluation Framework
• Preference Data Alignment and Ranking
• FAISS-based Corpus Indexing