Clash Royale AI — consolidated research artifacts
Code and navigation: https://github.com/cochon123/clash-royale-ai
This snapshot preserves all locally available model versions under models/,
including checkpoints, resumable runs, evaluation arrays and provenance archives.
reports/ contains HTML and supporting evidence; logs/ contains available logs.
snapshot_manifest.json records source Git revision and SHA-256 for each file.
The latest export is models/policy_v9/final_model/. It is experimental and
not promoted. Final card accuracy: 56.23%; joint card-and-cell accuracy: 12.76%.
The +1.55-point joint gain missed the required +2-point gate. Additional seeds
share a donor. These are offline metrics, not measured game win rates.
Read the export card, manifest and reports/policy_v9.html before using it.
The consumed final holdout is preserved as evidence, not a model-selection set.
hf download Cochon123/clash-royale-ai-artifacts --local-dir . \
--include 'models/**' 'reports/**' 'logs/**' 'snapshot_manifest.json'
Install the linked Python project and PyTorch for inference. Checkpoint/pickle files require the matching project code and trusted deserialization. The data snapshot is at https://huggingface.co/datasets/Cochon123/clash-royale-replays/tree/main/snapshots .