May 2026 Update from UT internal metamon

[jakegrigsby]

A large update that moves metamon development back to the public repo

Datasets + Showdown Version

• Everything has been synced up to May 20th 2026. Replays, usage stats, "revealed teams" all current as of May 19th.
• New self-play dataset "pac-tauros" focusing on high-GXE Gen1OU play
• New team sets built from recent replays and improved rating-dependent usage stat predictions. All legal as of today.
• We are back in sync with current pokemon showdown; metamon can run on public showdown despite recent breaking changes to sim mechanics with minimal drop in performance.

Models

• Lots of small Gen1OU specialist policies (V2A*) that were part of an ablation study. Released for extra opponent diversity in data generation
• TaurosV0 was recently rank 1 on PokéAgent Gen1OU. It is an effort to specialize on high-ladder Gen1OU, and contributed to the (still half-baked) test-time ensembling pipeline that let metamon hit rank 1 on the public showdown ladder pretty consistently in late April.
• New "grouped" observation space and associated TstepEncoder architecture.

Other

• Much improved pipeline (rl/evaluate) for launching local ladders for self-play, doing h2h evals, and generating learning curves.
• Model-based team prediction pipeline received a lot of internal work this spring but remains an experimental feature.
• Many other minor improvements