Good min/max clip scores for filtering for training #308

Hello, I'm trying to reproduce your training results using the train.py script with the same data.

I think using clip scores for filtering data is essential, as I found out some data cause trouble when training.

However, I couldn't find the min_clip_socre value and max_clip_score value used when training DALL-E mini/mega, and it's hard to reproduce the results without knowing the exact hyper parameters.

Could you share the value of clip scores you used for filtering data for DALL-E mini/mega? or did you not use the filtering feature when training your models?