ijcai-mahjong-ckpts-2026

IJCAI-2026 MCR Mahjong — Champion & Campaign Checkpoints

Model checkpoints from the IJCAI-2026 MCR (Mahjong Competition Rules) campaign. The competition entry is kdens3: a 3×KD-student mean-softmax ensemble. Code: https://github.com/SuuTTT/IJCAI-mahjong (see MANIFEST.md there).

Contents

champion/ — the DEPLOYED entry (kdens3)

• kdens_s{0,1,2}_fp16.npz — three distilled students. fp16 storage, fp32 compute, numpy-only (no torch at inference).
• numpy_resfused.py, feature.py — the inference stack used in the botzone bot.
• Load: build features with feature.py, run each kdens_s* via numpy_resfused.py, mean the softmaxes over the 3 students, argmax over legal moves.
• Metric: 2.5054 placement-pts vs aug_s0 baseline (ci_lo 2.5012; replicated 2.5057 / 2.5018) over ~200k paired identical-wall games. Sim-11: official 2nd of 25.

ckpt/kd/ — KD students (source of the ensemble)

• kd_128x40_s{0..5}.pkl + .bn.pkl — 128-wide × 40-block resnet students distilled from the aug teacher pool. .bn.pkl holds batchnorm running stats (needed for inference).

ckpt/aug/ — augmented BC baselines

• aug_128x40_s{0..13}.pkl + .bn.pkl. aug_s0 is the fixed A/B comparison anchor for every gate.

ckpt/value/ — value critic ensemble

• VALUE_C_60K.pt, VALUE_C_60K_s{1..7}.pt, value_e2e_ckpt.pt. Used in the E14 value-guidance study (genuine null result).

ckpt/placeval/ — placement + value heads

• placeval_s{0,1,2}.pt.

ckpt/oppbelief/ — opponent-belief models

• oppbelief_s{0..5}.pt, oppbelief_more60k_s{20,21,22}.pt (60k-step), oppbelief_big_s{10,11}.pt. .traininfo.json sidecars record hyperparameters.

ckpt/dealin/, ckpt/dealin_pc/, ckpt/dealin_pc_v2/ — deal-in predictors

• v1, pointcloud (pc), and pc_v2 deal-in probability nets used for the coherent-fold defense study.

ckpt/kdcurve/, ckpt/kd10/, ckpt/paperx/, ckpt/danger/ — ablation nets

• Teacher-count curve (kd1t/kd2t/kd4t), 10-teacher variant, paper ablations (a03/a09/half/quarter), danger/field-clone defense nets.

rl_league/<L>/latest.msgpack — representative RL league policies

• Final policies from selected league runs (Suphx-style GRP + oracle critic). Opponent pool.pkl files (~1.1G each) are not stored — regenerate during league training.

Training data

Corpora: Dannibal/ijcai-mahjong-corpora-2026. Eval testset: Dannibal/mcr-final2026-testset.