[Question] atari/montezumarevenge/expert-v0 only contains 0 returns trajectories

[Seraphli]

Question

I notice the dataset about Montezuma's revenge only contains 0 episodic reward trajectories. The document says the collecting script is using PPO Impala. Should this game use CleanRL PPO + RND to collect dataset?

https://wandb.ai/openrlbenchmark/openrlbenchmark/reports/-MontezumaRevenge-CleanRL-s-PPO-RND--VmlldzoyNTIyNjc5

[younik]

Hi @Seraphli,

For simplicity, all the Atari datasets where generated using PPO impala for all, with one script (https://github.com/Farama-Foundation/minari-dataset-generation-scripts/blob/main/scripts/atari/create_datasets.py). I am happy to add PPO + RND for Montezuma, as it make much sense.
Would you like to do it?

This means making a PR on https://github.com/Farama-Foundation/minari-dataset-generation-scripts, adding the script that generates the dataset; then I will host it on our remote.