Comment by indraneelpatil

Thanks! Its probably both, too little train data and insufficient quality.

Thats a cool idea, is there any VLM you would suggest? I can think of Gemini maybe? Or any would do?