Question about the Reinforcement Learning (RL) training data

[ZijunSong]

Hello,

Thank you for your great work on this project and for open-sourcing the model and the benchmark dataset.

I have a question regarding the training data. According to the paper, the model was trained in multiple stages, including reinforcement learning (RL) on specialist tasks and a final all-task RL phase.

I was wondering if you have any plans to also open-source the data used for these reinforcement learning training stages? Any information on this would be greatly appreciated.

Thank you!

[whitead]

Hi @ZijunSong,

At this time, we don't have plans to release all the training data. There are a few reasons for this, but one major one is we don't have licenses to re-distribute all data under a permissive license.

We do have a goal of releasing a subset of the data, probably chemical reaction data, to enable people to compare/repeat our training on that domain.

[zjlrock777]

@whitead Hi, when are you planning to release the chemical reaction data? Can you please also release the reaction data (retrosynthesis) in the benchmark so we can compare? Thanks