Datasets:
YGO-Turns (CoG 2026): request for preprint and reported baseline metrics
Dear Mr. da Silva Santos,
I am an MSc student in Artificial Intelligence and Cybersecurity at the University of Klagenfurt (AAU), Austria. For our Deep Learning course project, my group would like to use your YGO-Turns dataset for turn-level win prediction, and to compare our models against the results reported in your paper "YGO-Turns: A Large-Scale Tournament Dataset of Expert Yu-Gi-Oh! Duel States" (IEEE CoG 2026). We will of course cite the paper and the dataset.
Since we could not find a public version of the paper yet, I would be very grateful if you could share a preprint. In particular, it would help us to know:
- The reported metrics (AUROC, Brier score, and accuracy if available) for the logistic regression baseline, the scalar MLP, and the full WinPredictor.
- The exact evaluation protocol: the train/validation fraction, whether a separate test split was used, and whether the results are from a single run (seed 42) or averaged over several seeds.
- The source and version of the card data used to build cards.json, so that our card name matching and effect embeddings are as close as possible to yours.
Thank you very much for releasing the dataset and the code openly. It is a great resource for TCG research.
Kind regards,
Athanasios Pavlides
MSc AI & Cybersecurity, University of Klagenfurt (AAU)
Hi, thanks for your interest in the dataset!
The CoG 2026 proceedings haven't been published yet, so the paper isn't publicly available. If you leave me an email address, I can send you the IEEE paper directly, along with the cards.json file I used.
I also went back through the saved checkpoints to recover the Brier scores and accuracies. Here are the results:
| Model | AUROC | Brier score | Accuracy |
|---|---|---|---|
| Logistic Regression | 0.7469 | 0.2037 | 0.6756 |
| Scalar MLP | 0.7552 ± 0.0013 | 0.2033 ± 0.0007 | 0.6874 ± 0.0011 |
| BoardDSN | 0.7163 ± 0.0037 | 0.2151 ± 0.0014 | 0.6440 ± 0.0046 |
For the neural models, these are the mean and standard deviation across five runs (seeds 0–4), using the best checkpoint from each run.
The split is done at match level, with 80% for training and 20% for validation, using seed 42. I also noticed an error in Section IV of the paper, where I mention a test split. There was actually no separate test set in these experiments; the reported results are from the validation split. Sorry about that.
As for cards.json, I originally downloaded the card data from DuelingBook's static endpoint (https://static.duelingbook.com/cards.json) and removed some fields I didn't need. I didn't include my copy in the public repository because I was unsure about redistribution rights. I also don't have the exact DuelingBook version number anymore, but I still have the original file I used, so I can send you that by email as well.
Hope this helps, and feel free to reach out if you have any other questions!