TalBot

A small (~7.3M parameter) chess policy + value network that plays in the attacking style of Mikhail Tal, paired with a PUCT Monte-Carlo search so it can actually calculate.

• Strength: ~2070 Elo in a gauntlet against Stockfish (UCI_LimitStrength, Elo anchor 2000), 20 games, 1600 MCTS simulations/move, both colours.
• Style: predicts Tal's own move ~42% of the time on held-out games.

Architecture

ResNet: 18 input planes → 128 channels × 10 residual blocks →

• policy head → 4096 logits (from*64 + to),
• value head → tanh in [-1, 1] from the side to move.

Search: batched PUCT MCTS (policy priors + value leaves).

Training

1. Distillation from Stockfish (depth 8, MultiPV 4) on ~1.9M positions drawn from Lumbra's Gigabase (1950–1989) → general policy + value.
2. Style adaption: policy-head-only fine-tune (body and value frozen, lr 1e-4, 8 epochs) on 123k positions where Tal was to move.

stage	held-out policy top-1	value MSE
distillation	38.0%	0.055
+ Tal policy-head fine-tune	42.1% (Tal moves)	—

Usage

pip install torch python-chess numpy
# net = ResNet (see src/talbot/model.py); load model.pt (state dict + channels/blocks)
# move generation via MCTS: src/talbot/mcts.py (CompactMCTS)

Or run it as a UCI engine (drop it into any GUI):

PYTHONPATH=src python -m talbot.engine --model model.pt --sims 1600 --batch 24

Limitations

• Not a top engine. The Elo figure is anchored to Stockfish's limited-strength mode and measured over a small sample (20 games); treat it as ~2000 ± ~80.
• Style is a prior bias, not a reproduction of Tal's calculation. It prefers his openings and aggressive ideas, but cannot replicate his depth of sacrifice.
• The MCTS reference implementation is pure Python and is the main speed limit.

Data & licensing

• Games: public PGN collections (Lumbra's Gigabase; a curated Tal collection).
• Labels: generated with Stockfish (GPLv3) used only as an oracle; no Stockfish binary is distributed here.
• Code and weights: MIT.