Skip to content

notune/LeelaQueenOdds

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

25 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LeelaQueenOdds

This repo contains the network used on the LeelaQueenOdds Bot on Lichess, which you can download here lqo_v2.pb.gz.

Moreover, it also contains a customizable script that can be used to generate games to train your own odds-network. It supports various engine parameters, node limiting, opening books in FEN format, and automatic Dropbox file uploads. I also provided the configs and resulting training data used for networks v1 and v2.

Usage / Installation

Simplest way to play against LQO is on Lichess: https://lichess.org/@/LeelaQueenOdds

Disclaimer: Other use-cases such as analysis of normal chess positions, or even playing normal chess against the net is unlikely to work as intended as the net evaluates the queen-odds starting position as completely equal.

Debian Linux (Ubuntu, Linux Mint, ...)

Execute the following commands:

sudo apt update
sudo apt upgrade
sudo apt install ninja-build meson
git clone -b search-contempt --recurse-submodules https://github.com/Naphthalin/lc0.git
./lc0/build.sh
wget -P lc0/build/release/lc0 https://github.com/notune/LeelaQueenOdds/releases/download/v2/lqo_v2.pb.gz

the binary will then be located at lc0/build/release/lc0

Windows

If you have an NVIDIA GPU copy lc0_sc_cuda.exe and install-cuda_12_9.cmd into a new folder and double-click the install-cuda file.

Otherwise download lc0_win_sc_cpu.zip and unpack the file.

Then download the LeelaQueenOdds net: lqo_v2.pb.gz and copy it into the folder with lc0.exe.

Settings

Use the following settings when playing against the net: (The most important settings are ScLimit and SwapColors, ScLimit has to be set when searching for more then 1k nodes, and SwapColor is the only way to make it play as Black with reasonable strength).

When playing as Black (LQO plays White): Node limit (configure in chess gui): 15000

WeightsFile: lqo_v2.pb.gz
ScLimit: 40
CPuct: 1.5
FpuValue: 0.4
DrawScore: -0.4

When playing as White (LQO plays Black): Node limit (configure in chess gui): 12000

WeightsFile: lqo_v2.pb.gz,
SwapColors: true
ScLimit: 32
CPuct: 1.5
FpuValue: 0.4
DrawScore: 0.6

Training

Detailed instructions on how to train your own odds-network can be found here.

Acknowledgments

This project wouldn't be possible without the help from these people and projects:

  • The Leela Chess Zero Engine
  • Marcogio9 made the initial KnightOdds Network, inspired me to train my own queen-odds-net, and helped me a lot by providing me both the right engine configs to create the games, as well as the right training config, and answering any questions I had to be able to finish this project
  • Naphthalin and Hissha who answered any questions I had on the Leela Chess Zero discord server and helped me evaluate this network as well as find the right settings
  • Maia1900, the human-like chess network used to create training data for this net
  • Maia2200 from CallOn84 used for testing the network
  • python-chess used for generating games with the script
  • Chris Whittington for the Queen Odds opening book
  • amjshl implemented search contempt for lc0, which improves lqo's elo by another +200.
  • tiiber for sponsoring over 200$ in vast.ai compute for lqo_v3 (wip)

License

This project is licensed under the GNU AGPLv3 License, which you can read here: LICENSE.

About

Lc0 net trained specifically to win against humans with Queen Odds

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 2

  •  
  •  

Languages