Tutorial turns Tenhou mjlog logs into AI data, validated on 3,000 games

The post situates itself against earlier Japanese-language explanations of mjlog tags and meld codes, many written in Perl or built around conversion tools, and argues that what was missing is a single runnable Python path from raw log to training sample. It also stresses a validation habit: the replay function asserts hand consistency at each discard, and the author reports 3,000 games, 30,127 rounds and 1,441,017 discards with no errors. Those checks are presented as a way to catch parser mistakes rather than proof of correctness, and the author notes that mismatched tile types can still pass hand consistency, so cross-checks against things like yaku IDs are used as well.

The guide flags several traps that trip up would-be AI builders: seat order is relative to the starting dealer, tsumogiri (discarding the tile just drawn) has no explicit marker and must be inferred, double wins require storing multiple AGARI entries per round, and streamed hands omit melds. It warns that XML carries hidden information such as other players' hands and future dora, which, if fed to a model, would make it unusable in real play. The author limits the verified scope to 2023 four-player red-five Houou games and points readers toward shanten and tile-efficiency work as next steps.