NVILA reproducibility

[guangdongliang]

Good job!

Could I reproduce the model performance with your multi stage training code and data preparing code?

[Lyken17]

Sure, please check scripts/NVILA for stage 1,2,3 training.

For complete data list, please refer to paper datasheet section.

[guangdongliang]

Sure, please check scripts/NVILA for stage 1,2,3 training.

For complete data list, please refer to paper datasheet section.

Thank you very much!

[neelsj]

Can you please provide thew default.yaml and mixtures.yaml that go along with this. Right now these scripts refer to mixtures, such as llava_15_mix, that are not in the mixtures.yam file

[Lyken17]

Note we are not allowed to release the data directly. You have to download and prepare the data by your own.

[neelsj]

I understand this. I wasn't asking for the data, I'm okay to down and assemble it, but I was asking for the default.yaml and mixtures.yaml. So we can see exactly how the jsons are set up and if you are doing any subsampling, The data preperation code doesn't show this.