Laya RotomArena

What Laya RotomArena is, and whose teams taught it

Why it exists

Base Laya drafts the same team every time: Rotom, Torkoal, Torterra, Scrafty, Sinistcha and Gourgeist. Piloted by PokéChamp against the 228 teams the language models drafted, that team won 1 game in 60 and ranked last. A decision model has no way to look up what wins, so we showed it.

How it was taught

1. Drafts. 27 open-weight models drafted Pokémon Champions teams under c5’s own rules, prompts and referee: the teams they play on the ladder, plus 8 extra drafts each from the API models. 228 distinct legal teams in all.
2. A league of teams. PokéChamp piloted both sides of every game, so the teams were the only difference. Each team played at least 40 games against random others; a Bradley-Terry fit turned the results into ratings.
3. Lessons. Every team rated above the pool’s average became the questions Laya answers when it drafts, in the same wording and option lists, with that team’s choices as the answers: its six picks, then each Pokémon’s ability, item, nature, spread and four moves. Stronger teams count more. One team in ten was held out.
4. Keeping its battling. Laya’s own answers from 300 games against PokéChamp went in beside the drafting lessons, so the fine-tune keeps battling as base Laya did.
5. Training: two epochs with Laya’s own trainer (laya-train), on one consumer GPU, in about an hour.

What changed

It drafts far better and battles about as badly as before: it was taught to build teams, not to play them.

Whose drafts taught it

Every model below drafted teams in the pool; the number is how many of its teams rated above average and became lessons.

Ling 3.0 Tiny and Granite 4.2 8B drafted too, but none of their teams rated above average. Left out on purpose: every closed model (their terms forbid training on their outputs), MiniMax M3 (its license is non-commercial) and Drex (its license counts any model trained on its outputs as a derivative, and bars competing decision models).

The teams were rated with our PokéChamp, and Showdown’s validator refereed every draft, as on the ladder.