Aananda-giri/fly-chess

Fly Chess V6

Released under CC0 1.0 (public domain dedication) — use it for anything, no attribution required. That covers this repo's own weights, code and card only: the frozen connectome this checkpoint runs against is separately licensed CC BY 4.0 by its original publishers and is credited below; that attribution requirement travels with the connectome data itself, independent of the terms chosen for this derived checkpoint.

A small trainable "cortex" (~2.4M parameters) grafted onto a frozen, real MaleCNS fruit-fly connectome (166,606 neurons, from public HHMI Janelia / Google Research data, CC BY 4.0), trained to play chess. Chess features enter the model only through simulated fly sensory neurons; the move choice exits only through simulated fly motor neurons. At inference time, the network's opinion guides a classic alpha-beta search (up to depth 8, capture quiescence) rather than moving on its own — search always makes the final move selection, with the network contributing at most ±20 centipawns per move.

This is a research/demo artifact, not a general chess engine. It is bound to one fixed connectome and one fixed search configuration; unplugging either changes its behavior.

Release evaluation

Against pinned Stockfish 19, UCI_Elo 1320, 120+1, colors swapped, 18 of 30 planned opening pairs completed (36 games): release gate did not pass. W/D/L: 27/3/6. The benchmark stopped before every planned pair finished, so the gate could not pass on this evidence. Score 79%, an estimated 1552 on that Stockfish scale (95% lower bound 1300, no finite upper bound). This is a conditional Stockfish-UCI-Elo benchmark, not a FIDE, Lichess, or Chess.com rating. See v6/release.json for the full report, confidence interval, and exact checkpoint/engine hashes.

No search-only comparison (trained network switched off) was played, so this result does not show how much the network adds over search alone.

Source, training notebook, and the exact code that reconstructs this checkpoint's architecture are pinned alongside it in this repo (v6/model-source.ipynb, executed automatically by flychess/checkpoint.py's load_run). See the project repository for the full training/evaluation pipeline and a plain-language explanation of the design.

Files

• v6/manifest.json — exact architecture, training and search configuration
• v6/model-source.ipynb — pinned source reconstructing this checkpoint's code
• v6/main.best.pt — model weights
• v6/release.json — release benchmark result, if evaluated
• data/connectome/full-graph.npz, data/connectome/full-nodes.npz — the frozen connectome this checkpoint was trained against; a required input, not a training artifact. Inference will not reconstruct correctly without it.

Loading

from huggingface_hub import snapshot_download
from flychess.checkpoint import load_run
import chess

local = snapshot_download(repo_id="Aananda-giri/fly-chess")
ns = load_run(f"{local}/v6", f"{local}/data", f"{local}/data/connectome/full-graph.npz",
              device="cpu", checkpoint_name="main.best.pt")
move = ns["choose_v6_move"](chess.Board(), clock=120., increment=1.)