Today we are releasing the v5.3 “Quantum” dataset, which is ready immediately.
Features and Targets
The Quantum dataset contains 807 new features and maintains the same target list as before, but the new features improve model performance against these existing targets.
The payout target is remaining the same and we will continue supporting datasets v5.0, v5.1, and v5.2.
The following are correlation comparison results between Ender20 benchmark model predictions using V52 and V53 tournament datasets. It shows consistent improvements from V53 in correlations against the Ender20 target since 2018:
From now on this dataset is also available on Kaggle. You can add this dataset to your Kaggle notebook as a data source directly — either by selecting the published Kaggle dataset or by using the output of the data‑producing notebook. You don’t need to download the files yourself.
v5.2 live data will be supported indefinitely, but we will be consulting the data team about possibly de-commissioning train / val datasets for dataset v5-5.2.
Would it be possible to temporarily re-enable uploading v5.2 models on compute? At least until we have a few resolved rounds with the new v5.3 dataset.
