The repo contains the code for an AI media screening application. This service pulls data from the OME database, checks it with a simple prompt and gives a misinformation probability score. Scores >= 8 are flagged and saved on an object storage in order to be synced with labelstudio, that is here used as a journalist annotation platform.
The main logic of the service is deployed as a scaleway serverless job. You can find all serverless jobs in the jobs folder. The main logic is contained in the label-misinformation folder.
The project is templatised to be deployed for multiple countries, on multiple subjects. The OpenTofu code is found in the infrastructure folder. Refer to infrastructure/Readme.md for details.
The labelstudio interfaces are defined by xml files living in the deployments/labelstudio folder.
The notebooks used to create the dataset by pulling records from labelstudio and then analyse it live in the data folder
Scripts for finetuning models and for experimenting with different techiniques are in the climateguard folder under climateguard/finetuning and climateguard/experiments respectively.
Install dependencies with the setup.sh script.