This is not something you can train on a regular AWS gpu-instance without racking up millions of dollars in bills to my knowledge. Dataset isn't an issue its a capex issue.