ELO filtered/weighted transformer models

ELO filtered/weighted transformer models

• bidding transformer: filter on ELO > 950, weighted
• game eval transformer: filter on ELO > 950, weighted
• card play transformer: filter on ELO > 1100 (elite players), no game win filtering

Retrained bidding/cardplay transformers with architecture updates and tweaks

v1.3.0

Bump models release version

v1.2.0 - Fixed ONNX Data References

Fixes broken internal data file references in ONNX models.

Changes

• Re-exported all models with correct internal data file paths
• Removed card_play_dense (no checkpoint available)

Models

• bidding_dense / bidding_transformer
• game_eval_dense / game_eval_transformer
• card_play_transformer

Breaking Change

card_play_dense is no longer available. Use card_play_transformer instead.

v1.1.0 - Interface Specification

Adds interface specification files for implementing model consumers.

New in v1.1.0

• feature_spec.json - Card encoding, tensor shapes, input/output specs for all models
• test_vectors.json - Reference inputs and expected outputs for validation

Models (unchanged from v1.0.0)

• bidding_dense / bidding_transformer
• game_eval_dense / game_eval_transformer
• card_play_dense / card_play_transformer

v1.0.0 - Initial Release

Initial release of ONNX models for Skat ML.

Models included

• bidding_dense / bidding_transformer - Pre-skat bidding evaluation
• game_eval_dense / game_eval_transformer - Post-skat game type selection
• card_play_dense / card_play_transformer - Card play during tricks

Compatibility

• ONNX Runtime 1.19.2+