Bot trained on Jüri Ratas (currently Estonian Prime Minister) interviews and talks in Riigikogu (also from times when Jüri was not Prime Minister).
Bot is using transformer architecture. Data used for training model is not comprehensive. Mostly collected manually and using API for Riigikogu. Goal of this project is to have some fun and see if sequnece2sequence model could give human like answers. As Jüri is knwon for not giving direct answers some of the bot answers are not very different what you might when interviewing him.
To train model use notebook 0.1_prepare_jyri_data.ipynb to prepare data and notebook 0.2_train_jyri.ipynb to train model. Use GPU to seed up training. NB! in requirements.txt there is only CPU version of pytorch, you have to install GPU version for training. Model training was done using python 3.6.6.
To use bot have model files in models sub folder. Use build_dev.bat (you need to convert it.sh files if using linux) to build docker image. Use run_dev.bat to run docker container. Open browser and go to localhost:5001 (or 127.0.0.1:5001) and use.
Thanks Samuel Lynn-Evans for transformer tutorial and code. I've slightly modified code and turned it to dockerised flask app.