Can LLMs Beat BERT in Biomedical Information Extraction? Evaluating Prompting and Fine-Tuning Strategies for NER and Classification

Author: Vera Bernhard
Date: December 2025
Institution: University of Zurich, Switzerland

---

This repository contains the code and data for the Master’s thesis by Vera Bernhard.

Structure

• bert_baseline/: Prediction files and evaluation outputs for the BERT baseline models
• data/: The PsyNamic dataset
• evaluation/: Evaluation, post-processing, and plotting scripts
• few_shot/: Predictions and plots for the few-shot experiments
• finetuning/: All files related to fine-tuning LLMs
  • ift/: Instruction fine-tuning dataset and training scripts
  • lst/: Label-supervised fine-tuning scripts and predictions
• prompts/: Prompt templates, prompt generation scripts, and annotation guidelines for the PsyNamic dataset
• test/: Unit tests for evaluation and post-processing scripts
• zero_shot/: Predictions and plots for zero-shot experiments, including predictions from the instruction fine-tuned model

Technologies Used

• Python 3.12
• Hugging Face Transformers – model loading, inference, and training
• PEFT – parameter-efficient fine-tuning methods
• TRL – training large language models with instruction tuning
• BiLLM – converting LLMs from uni-directional to bidirectional for classification tasks