Details of training script

[dreminm]

Hello!

I'm currently trying to understand the model training process and noticed the use of the --data_path ./llava/state3_after.json parameter. Could you provide more details of this file? As far as i understand this script provides stage3 finetuning. Do you have script for pretraining you model with distortion classification and score classification (stage 1+2)?

[jzhws]

The three stages are trained with the same script with only a little adjustments. So you can use this code to finish the training of the three stages. Besides, the stage-1 data is set to be private so you may only train the stage-2 and stage-3.