Open science

We release the code behind our published work, with instructions to reproduce the main results.

Where licensing and privacy allow, we release datasets and document how they were collected and processed.

We release trained models when doing so is safe and useful, with clear documentation of their limits.

We report seeds, splits, hardware, and hyperparameters, and we publish reproducibility reports on our own work.

We compare against strong baselines on identical splits and report where our methods fail.

We follow the data use agreements of every dataset we use and never release data we are not permitted to share.