Good Gyal: a different kind of Leela network

What if you could combine the smarts of stockfish with the sleazy and swashbuckling play of lichess humans? Good Gyal (Jamaican for “good girl”) does just that. It’s taken 1.3 million lichess games and turned them into Leela Chess training data with the help of stockfish. It plays a very different game from self-play trained leela nets.

github.com/dkappe/leela-chess-weights/wiki/Bad-Gyal

You can play against it on lichess as MiniHuman. It runs with 8 nodes per move on a raspberry pi 3 (the network is a tiny 48x5). For added fun, it uses a gambit opening book.

Would appreciate comments and feedback.

No feedback? Fun to play against? Good training partner?

It’s gratifying seeing more people playing against Good Gyal. But I’m curious whether people find it fun/useful to play against her.

Next version will include 48x5 Evil Gyal (0.25 q-ratio), 128x10 Bad Gyal (0.5 q-ratio) and 192x16 Good Gyal (0.75 q-ratio). Evil Gyal will take over playing under MiniHuman.

Hi there, I've played against Minihuman and I founded very fun...
keep going and best regards :-)