Hallo Dietmar,
Hier noch folgende Info:
It is built on a Leela Chess Zero BT4 transformer body with a Stockfish-aware policy head.
chessard:
A UCI chess engine that plays like a human of a chosen rating. Instead of searching for the best move, it plays the move a person of that strength would most likely play, as predicted by the chessard network.
Quick start
Works on Windows, macOS and Linux. You need Git, Python 3.10 or newer, and about 1 GB of disk space.
Windows (PowerShell or Command Prompt; when installing Python, tick "Add python.exe to PATH"):
git clone
https://github.com/daniel-monroe/chessardcd chessard
py install.py
venv\Scripts\chessard.exe
macOS / Linux:
git clone
https://github.com/daniel-monroe/chessardcd chessard
python3 install.py
venv/bin/chessard
install.py sets everything up inside the chessard folder, then runs verify.py to check that the engine works. It takes a minute or two. Re-running it only redoes what is missing.
Step What it does Ends up in
Python a virtualenv with the pinned versions in requirements.txt, plus PyTorch 2.14.1 (2.2.2 on Intel Macs, the last release for them) venv/
Engine command chessard (chessard.exe on Windows), a normal program to give to any chess GUI venv/bin/ or venv\Scripts\
Stockfish the official Stockfish 18 release for your OS and CPU. On Linux, if no release runs (an old glibc, or ARM), it is built from source, which needs make and g++ stockfish/
Weights downloaded from Hugging Face (public, ~400 MB, no login) and checked against weights.sha256 weights/
Options:
--cpu / --cuda: choose the PyTorch build on Windows and Linux. The default is CUDA when an NVIDIA GPU is found (nvidia-smi works); it's a ~2.5 GB download, against ~200 MB for CPU. Macs always get the standard macOS build.
--weights-src DIR: copy the weights from a local folder instead of downloading them.
--skip-verify: don't run verify.py at the end.
verify.py checks that:
the UCI handshake lists all five players;
at Elo 2200 after 1.e4 d5, the first info line is score cp 91 ... pv e4d5 string p=93.81% and the move is bestmove e4d5;
--player kaufman loads at Elo 2188;
--elo 1500 is rejected.
Every push is tested this way on Windows, macOS and Linux (.github/workflows/test.yml).
Running it
venv/bin/chessard # Elo 2200 (Windows: venv\Scripts\chessard.exe)
venv/bin/chessard --elo 2400
venv/bin/chessard --player carlsen # Carlsen's style, at his rating (2840)
The engine finds weights/ and stockfish/ in the repo on its own. To use other copies, pass --weights-dir / --stockfish or set CHESSARD_DIR / STOCKFISH. The weights folder can be shared between programs. The Player option lists every <name>.pt in its loras/ subfolder, so an adapter you train yourself shows up once you copy it in; add it to loras/players.json to give it a default rating.
GUIs and lichess-bot
Give the GUI the full path to the engine command: C:\...\chessard\venv\Scripts\chessard.exe on Windows, /.../chessard/venv/bin/chessard on macOS and Linux. install.py prints it at the end.
Cute Chess, Arena, BanksiaGUI, En Croissant, Nibbler: add a UCI engine with that path, then set Elo, Player, Sampling and so on in the engine options dialog.
cutechess-cli / fastchess:
cutechess-cli -engine cmd=/path/to/chessard/venv/bin/chessard name=chessard-2400 option.Elo=2400 \
-engine cmd=stockfish option.UCI_LimitStrength=true option.UCI_Elo=2400 \
-each proto=uci tc=60+1 -games 2
lichess-bot (config.yml):
engine:
dir: "/path/to/chessard/venv/bin/" # Windows: C:\path\to\chessard\venv\Scripts\
name: "chessard" # Windows: chessard.exe
protocol: "uci"
uci_options:
Elo: 2400 # 2000-2900, or 0 for the player's own rating
Player: "none" # or carlsen, nakamura, sadler, janik, kaufman
Options
Option Default
Elo 0 (auto) rating to imitate, 2000–2900: the model was trained only on games by 2000+ players, and below that its behaviour is not meaningful. 0 means the selected player's own rating from loras/players.json, or 2200 with no player
Player none play in one player's style using their adapter: carlsen (2840), nakamura (2810), sadler (2692), janik (2504), kaufman (2188)
Sampling true sample a move from the distribution instead of always playing the most likely one; gives varied, more human games. --no-sampling turns it off
Temperature 50 percent; 100 is the network's own distribution, >100 flattens it, <100 sharpens it
MultiPV 5 how many candidate moves to report
SF_Depth 9 depth of the per-move Stockfish searches. The network was trained on depth 9; other depths work but shift its predictions
Threads min(8, cores) Stockfish processes run in parallel
Move Overhead 100 ms reserved for GUI/network lag
WeightsDir, StockfishPath same as --weights-dir, --stockfish
Set the CHESSARD_DEVICE environment variable (cpu, cuda, cuda:1) to choose a device.
Time and stop
chessard does not search, so extra time doesn't make it better: each move costs one network forward pass plus a shallow Stockfish search of every legal move. The clock only acts as a safety cap (25% of remaining time plus most of the increment). If that runs out, or stop arrives, before the searches finish, the move is chosen from depth-1 evaluations instead, which is much cheaper. go infinite and go ponder hold bestmove until stop / ponderhit, as UCI requires. searchmoves is supported; depth, nodes and mate limits are accepted and ignored.
Troubleshooting
python/py is not found (Windows). Reinstall Python from python.org and tick "Add python.exe to PATH", or run it through the py launcher, which the installer adds.
could not create a virtualenv (Debian/Ubuntu). Run sudo apt install python3-venv.
The weights download fails. The repo is public, so this is usually the network (a proxy, a firewall, or an interrupted transfer). Re-run install.py and it resumes. Behind a proxy, set HTTPS_PROXY. You can also fetch the files another way and pass --weights-src DIR.
the weights don't match weights.sha256. A download was corrupted. Delete weights/ and re-run.
No Stockfish release runs, and the build needs make. This happens on Linux with an old glibc, or on ARM. Install a compiler (sudo apt install build-essential) and re-run.
Intel Mac: "PyTorch only goes up to Python 3.12". PyTorch's last Intel Mac release (2.2.2) supports Python 3.10–3.12, so run the installer with one of those (e.g. python3.12 install.py).
The GUI shows no moves, or chessard failed to load. Run the engine command in a terminal and type uci, then isready; errors go to stderr. The usual causes are a wrong WeightsDir/CHESSARD_DIR or StockfishPath.
--elo 1500 is rejected. This is intentional, because the model was trained only on games by 2000+ players. Through setoption, out-of-range values are clamped to 2000-2900.
It runs on CPU even though there's an NVIDIA GPU. You have the CPU build of PyTorch. Re-run install.py --cuda to swap it.
verify.py fails only on the probability. It allows ±0.05 points. A bigger difference means a different Stockfish version, SF_Depth or weights.
Starting over. Delete venv/, weights/ and stockfish/ and run install.py again.
Python API
import chess
from chessard import Chessard
net = Chessard("chessard-weights/chessard.pt", stockfish="stockfish",
adapter="chessard-weights/loras/carlsen.pt") # adapter is optional
for m in net.predict(chess.STARTING_FEN, ["e2e4", "e7e5"], elo=2200)[:3]:
print(m["uci"], f"{m['prob']:.1%}", m["cp"])
net.close()
Pass the real move sequence rather than a bare FEN. The network sees the last 8 positions, so predictions from a FEN with no history are worse.
License
GPL-3.0. See LICENSE.
Forum
CSS-Shop