chess-transformer-100m-overnight_20260908

Chess transformer 100M — overnight SF19 SWA (2026-09-08)

Searchless greedy policy. Incumbent after a matched 128-game SF19 2050/2200 screen against two 8k fine-tunes. Those mixes did not beat this checkpoint on a paired-opening bootstrap.

This is not FIDE or Lichess Elo. Local UCI_Elo screen ≈ 2150 (2050 score 0.602, 2200 score 0.445, 128 games, 8000 nodes, no book / no Syzygy). A same-protocol re-run scored 0.504 overall (estimate 2132). Treat the band, not a single number.

The public repo avewright/chess-transformer-100m-squares64 was not overwritten.

Files

• eval_swa.pt / latest.pt — evaluation weights (same SWA snapshot)
• training_resume.pt — full overnight live checkpoint (optimizer + RNG)
• model_config.json — Squares64RecurrentConfig
• evaluation/ — gauntlet, holdouts, mix audits

Inference

import os
os.environ["MOVE_VOCAB_VERSION"] = "compact"
import chess
from huggingface_hub import hf_hub_download
from chess_inference import load_checkpoint, get_model_move

path = hf_hub_download("avewright/chess-transformer-100m-overnight_20260908", "eval_swa.pt")
model = load_checkpoint(path, device="cuda")
move, info = get_model_move(model, chess.Board(), "cuda")

Compact vocab 1968. No search at inference.