Welcome to the repository for the thesis titled "Application of Machine-Learning Algorithms to Detect AI-generated Content".
This repository contains the essential files related to the thesis.
This directory contains all the data sets that are used to train the models. While the data sets that have been created newly are in a dedicated directory, the provided data sets are at the top level.
This repository contains all the datasets described above. responsesLlama.csv and responsesMistral.csv had been created by myself using capacities of the university.
This directory contains all the measurements used to analyze the models. Graphics and charts that are cited in the written part from this repository can be found here.
Contains measurements about the Cross Validation procedure.
Contains measurements about the GPT 3.5 dataset.
Contains measurements about the GPT Neo dataset.
Contains measurements about the Llama 3 dataset.
Contains measurements about the Mistral dataset.
In this directory, the Large-Language-Models Llama3 and Mistral are deployed and executed. When executing the run files, it should be ensured that a sufficient graphics card is available, as the code is optimised for this and these models do not actually run without a graphics card. The datasets that are created are stored directly in the datasets directory.
Contains the code for executing the Llama3 model locally.
Contains the code for executing the Mistral model locally.
The model directory contains the implementation of each classifier. Each classifier is applied on each dataset individually as well as the ensemble models. The evaluation metrics are stored directly in the evaluation directory. The ensemble models just load the classifiers from the pickle directory, therefore, no additional computation is needed to run them.
Contains the Cross-Validation-models for each data set. This means each pre-trained model is applied on each data-set. The particularity here is that each pre-trained model has been trained on 100% of the available training data.
Contains the ensemble-models for each data set.
Contains the logistic regression models for each data set.
This directory contains all of the executed models and saves them by model and data set. Important: The saved files are overwritten if a model file is executed.
Contains the random forest models for each data set.
Contains the support vector machine models for each data set.
This directory contains the files to preprocess each data-set. Each prepared data-set is stored in the storePreprocess subdirectory.
DOI:
