jjeyxb/mia-yolo

MIA YOLO weights

Screen-recognition models used by MIA (Mahjong Intelligence Assistant) for Mahjong Soul: game regions, discard ponds, the player's hand, and called melds.

They are loaded by the YOLO worker in engines/yolo/ of that repository and are bundled into MIA's installers. Developed by a MIA team member and published here with the author's permission.

Files

File	Size	SHA-256
hand_model_best.pt	5.9 MB	04a47c872baea0f548e4c473e80d6e685f06e31775ac375f016f79c2cd87ea94
meld_model_v32_1280.crop_meta.json	0.0 MB	c0e4752a2045f0529a7fcf38a2db8153a571925e8236248232676c24b762dd87
meld_model_v32_1280.infer.json	0.0 MB	73f83759cf5a3368ce5a7eda25fe8b6d032764f5720edfd73b0cd475f0039431
meld_model_v32_1280.pt	5.2 MB	742f0720c64abaa095a66d3b0035d329b05c8dcbbde3f40184721c4c8f6fba38
region_model_best.pt	19.4 MB	160570dbd1e3f9977dd80267076622d59821e27447ab4194d7e5af2595c0ff68
river_model_best.infer.json	0.0 MB	73f83759cf5a3368ce5a7eda25fe8b6d032764f5720edfd73b0cd475f0039431
river_model_best.pt	19.4 MB	54b5cb9eacc8361541953c1edc445d35767c1490571a1af4ff41cbae4020e367

The .json files are inference and crop settings read next to the matching .pt file.

Usage

From a MIA checkout:

python tools/fetch_models.py

This downloads these files into models/yolo/ together with the Mortal weights.

License

AGPL-3.0. The models are trained with Ultralytics, whose models are distributed under AGPL-3.0, consistent with MIA itself.