Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -15,7 +15,7 @@ datasets:
|
|
| 15 |
# MTG Commander Deck Completion
|
| 16 |
|
| 17 |
Given a Commander (and optionally a partial decklist), suggests which other
|
| 18 |
-
cards belong in the deck. Trained on
|
| 19 |
from [Moxfield](https://moxfield.com) ([dataset](https://huggingface.co/datasets/nsaroiu/moxfield-dump)).
|
| 20 |
|
| 21 |
Training/evaluation code: [github.com/nsaroiu/mtg-deck-completion](https://github.com/nsaroiu/mtg-deck-completion)
|
|
@@ -74,13 +74,13 @@ empty), staple-recognition checkpoint, validation split:
|
|
| 74 |
|
| 75 |
| mask ratio | Recall@50 | Precision@50 | MRR |
|
| 76 |
|---|---|---|---|
|
| 77 |
-
| 0.1 (deck mostly complete) | 0.
|
| 78 |
-
| 0.5 | 0.
|
| 79 |
-
| 0.9 (deck mostly empty) | 0.
|
| 80 |
|
| 81 |
Held-out-commander split (commanders never seen in training, testing
|
| 82 |
generalization via card content rather than memorized co-occurrence):
|
| 83 |
-
Recall@50 0.
|
| 84 |
confirming the hybrid content embeddings carry real signal for unfamiliar
|
| 85 |
commanders. Full breakdown by card-popularity tier, plus the synergy/pruning
|
| 86 |
checkpoints' own numbers, in the GitHub repo's eval output.
|
|
@@ -151,7 +151,8 @@ fusion, if that's all you need.
|
|
| 151 |
## License
|
| 152 |
|
| 153 |
The weights and code in this repository are released under the MIT license.
|
| 154 |
-
The training data itself is a scrape of public Moxfield decklists
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
a constraint this repo
|
|
|
|
|
|
| 15 |
# MTG Commander Deck Completion
|
| 16 |
|
| 17 |
Given a Commander (and optionally a partial decklist), suggests which other
|
| 18 |
+
cards belong in the deck. Trained on 189,108 real public Commander decklists
|
| 19 |
from [Moxfield](https://moxfield.com) ([dataset](https://huggingface.co/datasets/nsaroiu/moxfield-dump)).
|
| 20 |
|
| 21 |
Training/evaluation code: [github.com/nsaroiu/mtg-deck-completion](https://github.com/nsaroiu/mtg-deck-completion)
|
|
|
|
| 74 |
|
| 75 |
| mask ratio | Recall@50 | Precision@50 | MRR |
|
| 76 |
|---|---|---|---|
|
| 77 |
+
| 0.1 (deck mostly complete) | 0.57 | 0.10 | 0.121 |
|
| 78 |
+
| 0.5 | 0.48 | 0.40 | 0.066 |
|
| 79 |
+
| 0.9 (deck mostly empty) | 0.36 | 0.54 | 0.045 |
|
| 80 |
|
| 81 |
Held-out-commander split (commanders never seen in training, testing
|
| 82 |
generalization via card content rather than memorized co-occurrence):
|
| 83 |
+
Recall@50 0.29–0.44 across the same ratio range — meaningfully above chance,
|
| 84 |
confirming the hybrid content embeddings carry real signal for unfamiliar
|
| 85 |
commanders. Full breakdown by card-popularity tier, plus the synergy/pruning
|
| 86 |
checkpoints' own numbers, in the GitHub repo's eval output.
|
|
|
|
| 151 |
## License
|
| 152 |
|
| 153 |
The weights and code in this repository are released under the MIT license.
|
| 154 |
+
The training data itself is a scrape of public Moxfield decklists (see the
|
| 155 |
+
[dataset repo](https://huggingface.co/datasets/nsaroiu/moxfield-dump), now
|
| 156 |
+
gated/manually-approved) with no asserted license — if that matters for
|
| 157 |
+
your use case, that's Moxfield's terms to check, not a constraint this repo
|
| 158 |
+
imposes.
|