Documentation for Go Experiments and VideoGobench

[outerform]

Hi,

I am trying to reproduce the results from this work, but I couldn't find instructions for training and evaluating the Go experiments. Additionally, there doesn't seem to be a README file for VideoGobench. Could you please provide this information?

Thank you!

[MaverickRen]

In the LDM part, the training process for Go is the same as that for CALVIN. During the second phase of autoregressive training, the evaluation of Go relies on KataGo's scoring. Specifically,

VideoWorld/VideoWorld/falcon/models/algorithms/go_battle_with_human.py

Line 423 in 2e1efee

record = {'eval_dict': eval_dict}

saves KataGo's scores in Eva_dict, where "prior_ratio_pred_to_katago" corresponds to the Action-Value mentioned in the paper.

[MaverickRen]

Regarding data usage, since the VideoGoBench dataset is very large, we employ WebDataset for data loading. The dataset is stored on Huggingface in tar-based format.

For testing purposes, a small subset of the data can be used, eliminating the need for WebDataset-based loading.