YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
IJCAI-2026 Chinese-Standard-Mahjong — Phase-1 artifacts
Models and datasets from the IJCAI-2026 Botzone Chinese-Standard-Mahjong campaign.
Code + write-up: https://github.com/SuuTTT/IJCAI-mahjong · Autopsy: docs/phase1_autopsy.html.
models/
| file | what |
|---|---|
cnn_lad_chunjiandu.pkl / .npz |
The lock / submission. 40-block ResNet (128ch) distilled from the strongest coherent ladder teacher. |
cnn_distill100b.pkl |
Safe-floor multi-teacher BC baseline. |
strong5_aw_b0.3.pkl, strong5_aw_b0.5.pkl |
Cycle-1 strong-teacher distill (8,888 dec, AWBC). |
strong5_full_aw_b0.3.pkl |
Cycle-2 pooled strong-5 (24,401 dec). |
typec_full_aw_b0.3.pkl, typec_full_aw_b0.5.pkl |
Cycle-2 TypeC青雀 single coherent teacher (7,733 dec). |
mythos_full_aw_b0.4.pkl |
Cycle-2 mythos (strongest bot, 4,104 dec). |
Result: none of the distill candidates beat lad_chunjiandu in 144-game duplicate gauntlet
(best, TypeC β0.3, ties at −24/144g). lad_chunjiandu stays the submission. See repo paper/evidence/.
data/
Strong-teacher decision datasets (obs, mask, act, score for advantage-weighted BC), extracted
from Simulation-8 duplicate logs of the field's strongest non-LLM bots.
| file | decisions |
|---|---|
strong5.npz |
8,888 (pooled, opponent-appearances) |
strong5_full.npz |
24,401 (pooled, full target sets) |
typec_full.npz |
7,733 (TypeC青雀, SIM-8 + global) |
mythos_full.npz |
4,104 (mythos) |
Format matches deploy/caiest_cnn/feature.py (38-plane obs, 235-action space; Play=indices 2..35).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support