avaskys/skat-ml-models

Skat ML Models

Models for playing the German card game Skat, trained on ~9 million games from the International Skat Server (ISS).

Models

Model	Architecture	Description	Size
bidding_dense	Dense ResNet	Pre-skat bidding evaluation	~9 MB
bidding_transformer	Transformer	Pre-skat bidding with attention	~14 MB
game_eval_dense	Dense ResNet	Post-skat game type selection	~9 MB
game_eval_transformer	Transformer	Post-skat evaluation with attention	~13 MB
card_play_transformer	Transformer	Card play selection	~72 MB

Gameplay Phases

Skat has three decision phases, each with its own model:

• Bidding — "Can I win at bid level X?" Evaluates win probability across all 63 bid levels.
• Game selection — "Which discards + game type maximize win probability?" Evaluates after winning the bid.
• Card play — "Which card should I play?" Selects from legal cards during tricks.

Model Details

Model	Input	Output
BiddingEvaluator	35 features (10 cards + position)	63 × 2 win probs (pickup/hand)
BiddingTransformer	10 card indices + position	63 × 2 win probs (pickup/hand)
GameEvaluator	75 features (hand + skat + context)	1 win probability
GameEvaluatorTransformer	Card indices + context	1 win probability
CardPlayTransformer	Sequence of moves	32 card logits

Interface Specification

The interface/ directory defines the contract for using these models:

• feature_spec.json — Card encoding, game type encoding, input/output tensor shapes
• test_vectors.json — Reference inputs and expected outputs for validation

Use these to ensure your implementation encodes inputs correctly.

Usage

Load with ONNX Runtime in any supported language (Python, Java, C++, etc.):

import onnxruntime as ort

session = ort.InferenceSession("bidding_transformer.onnx")

Training

Models were trained with rating-weighted loss using ISS player ELO ratings, so decisions from stronger players contribute more to the learned policy.

For training code and details, see the GitHub repository.

License

MIT