Could you link to your research and/or describe the models, libraries, data and tests you used for this?