Persona-Conditioned Clinical Reasoning: Measuring Country-Cue Sensitivity in Large Language Models Using Synthetic, Clinician-Vetted Diagnostic Vignettes

Overview

This repository contains materials for a research project that evaluates whether a minimal geographic persona cue in the system prompt (e.g., "You are a physician in {COUNTRY}") affects large language models' diagnostic reasoning.

The conceptual experiment utilized multiple models, diverse persona conditions, 13 clinical vignettes, and a rubric-based grading approach. This README documents the repository's current, runnable contents and how they map to the original conceptual workflow.

Current Status & Intent

• Primary Entry Point: The repository contains a Jupyter notebook, datasets (CSV), and per-model output folders.
• Structure: It does not currently contain separate, modular scripts for generation, grading, and analysis (i.e., the original generation/, grading/, analysis/, and results/ directories are not present).
• Usage: The notebook contains the core code and analysis steps; consider it the central tool for reproducing or adapting experiments.

Repository Contents

├── code/
│   └── diagnostic_vignettes.ipynb   # Main notebook (Data loading, prompts, analysis, viz)
│
├── dataset/
│   ├── diagnostic_vignettes.csv     # The 13 clinical vignettes (CSV)
│   └── rubrics.csv                  # Scoring rubrics used for grading (CSV)
│
├── outputs_by_models/               # Storage for model-specific outputs
│   ├── deepseek-chat/
│   ├── gpt4omini/
│   ├── gptoss20B/
│   ├── llama8b/
│   ├── mistrallarge2512/
│   └── mistralsmall2506/
│   # Note: Subfolders hold model-specific outputs; some may be partially populated.
│
└── README.md