CTPL: Contrastive Trajectory-Aligned Preference Learning for Mental Health AI

This repository contains the implementation code, experimental scripts, configuration files, and reproducibility materials for the study on AI-driven mental health preference modelling and therapeutic response evaluation. The project investigates how machine learning and language-model-based representations can be used to model conversational response preferences, estimate therapeutic alignment, and evaluate mental-health-oriented dialogue behaviour using structured experimental pipelines. The codebase includes scripts for data preprocessing, preference-pair construction, embedding generation, baseline model training, trajectory-based evaluation, cross-validation experiments, and high-harm response coverage analysis. It also contains experiment outputs and configuration files used to reproduce the reported results, including trajectory correlation, cross-validation correlation, and high-harm coverage metrics. The main components of the repository are: - Data preprocessing and cleaning scripts- Preference-pair and trajectory label construction utilities- Embedding and representation generation modules- Baseline model training and evaluation scripts- Hyperparameter and configuration files- Cross-validation and robustness evaluation scripts- Experimental result files for reproducibility- LaTeX manuscript files associated with the research paper This code release is intended to support transparency, reproducibility, and further research in AI-assisted mental health dialogue modelling, preference learning, and safe conversational AI. The repository may be useful for researchers working on mental health NLP, reinforcement learning from human feedback, therapeutic dialogue systems, explainable AI, and safety-aware language model evaluation. Please cite this Zenodo archive and the associated paper if you use this code or build upon it in your own research.