Add all necessary files to replicate training run

Use 1 caption per image

Update downloading and training scripts

Add training scripts and initial model trained on 1% of the data.