yoheikobashi/ptcg-dusknoir-deberta-reranker

PTCG dragapult_dusknoir DeBERTa reranker (fld_r49b)

Cross-encoder move reranker for The Pokemon Company "PTCG AI Battle Challenge" Kaggle simulation (2026). Pilots the dragapult_dusknoir deck: the game state is serialized to a compact text prompt (no glossary, board + hidden facts, opponent-deck ID) and each legal move is scored as a (state, candidate) pair; argmax plays.

• Backbone: DeBERTa-v3-base + ~3k domain tokens (card/attack/deck ids), fp32-trained.
• Training: SFT on engine self-play, then ~67 rounds of mirror-RL / field DPO on playout-gated pairs from the policy's own low-margin decisions, gated per round against an 8-deck field (ship on positive delta only). This checkpoint (round 49, arm b) defended the gate for the final 18 rounds.
• Deployed as submission dusk_v3/v4: vocab-pruned + weight-only INT8 ONNX (in onnx/), 166 MiB bundle, 2 vCPU, 480 s time bank with engine fallback. Final live rating 398.2.
• Final local field gate (150 games/opponent, engine_v2 opponents): 43.9% overall -- ethan_hooh 63.3 / dudunsparce 62.0 / dragapult 59.3 / marnie 34.7 / alakazam_nz 28.7 / ogerpon_mono 18.7.

Code: https://github.com/yohei-kobashi/pokemon-card-bert (see lm/, tools/, tools/instance/).