daydream-chess-nanogpt-grand-1

Model Card — daydream-chess-nanogpt-grand-1 (v1, Grand)

A sup computer release — a small language model studio. Model page · monorepo (frozen code: projects/daydream/models/daydream-chess-nanogpt-grand-1/, tag daydream-chess-nanogpt-grand-1) · runs in your browser at www.supcpu.com/model-player.

<div class="takeaways"> <p class="takeaways-label">Key takeaways</p> <ul> <li>A 4.73M-param char-level GPT trained on a custom 12-file × 10-rank chess variant with Chancellor and Archbishop pieces added — the largest, most complex board in the <a href="../../projects/daydream/README.md">daydream</a> family.</li> <li>The board spec was <strong>corrected mid-build against the live engine</strong>, not trusted from research: real Grand Chess has one Chancellor and one Archbishop per side (not two, as an earlier web search claimed), and the installed Fairy-Stockfish caps board ranks at 10 even though files go to 12 — so this is 12×10, not 12×12.</li> <li>100% clean completion, 36.7% legal-move rate on first try — in line with Regular's 35.3% and Micro's 39.2%, a <strong>consistent legality-learning signal</strong> across all three board sizes despite very different vocabularies and piece sets.</li> <li>Promotion needed <strong>zero new tokenizer characters</strong>: Capablanca-style six-way promotion (Q/Chancellor/Archbishop/R/B/N) is fully covered by letters already in the 12-file alphabet (a–l) plus n/q/r, landing on a 27-character vocabulary.</li> </ul> </div>

The largest, most structurally novel tier in the daydream family: a chess-move GPT trained on a custom variant extending real Grand Chess to 12 files × 10 ranks, with two Chancellor (rook+knight) and two Archbishop (bishop+knight) pieces per side. Same mechanic as the rest of the series: legal moves snap into focus, illegal moves render as dim near-misses instead of being discarded.

A wider board needs denser pieces, not just more empty squares. Historical large-board chess variants (Capablanca Chess, real Grand Chess) don't just spread the standard 16 pieces over more squares — they add compound pieces to keep tactical density proportional to the board. Grand extrapolates that logic one step further than its 10×10 namesake, adding a second Chancellor and Archbishop pair to keep density proportional at 12 files.

Model details

Intended use

Same exhibit posture as Regular and Micro, at the largest board in the series. Pairs with harness.py (this folder), which plays the model against Fairy-Stockfish under the vendored custom grand12 variant.

Out of scope. Not a chess engine, not evaluated for playing strength. Vocabulary and board are grand12-specific — Chancellor and Archbishop moves and this board's square names are meaningless on Regular's or Micro's boards and vice versa.

Training data

No human corpus exists for this board, so it's entirely synthetic: 2,101 self-play games between two Fairy-Stockfish instances under the project's custom grand12 variant (vendored as variants.ini in this folder), bounded-depth search with randomized opening plies for diversity. Corpus vendored in-folder as games.txt (synthetic, seeded, code-owned — committed).

The board, exactly

Extends real Grand Chess (10×10) rather than inventing a new design from scratch. The base variant, confirmed by querying the live engine rather than trusted from documentation, has rooks alone in the back-rank corners, a mirror-asymmetric second rank with one Chancellor and one Archbishop sitting off-center next to the king, and pawns filling the entire third rank. Grand's 12×10 extension makes the second rank fully mirror-symmetric by adding a second Chancellor/Archbishop pair: Knight, Bishop, Chancellor, Archbishop, Queen, King, Archbishop, Chancellor, Bishop, Knight across files b–k (a and l stay empty, matching the base variant's corners), with rooks at a1/l1 and a full pawn rank at rank 3. 24 non-pawn pieces plus 12 pawns per side.

Training procedure

• Optimizer: AdamW, LR 3e-4 with cosine decay to 3e-5, 100 warmup iters, β₂ 0.99, batch size 48.
• Run: 3,000 iterations, best val loss 1.053.
• Hardware: Apple Silicon Mac (MPS / Metal backend), torch.compile disabled.

Evaluation

Consistent with the other two tiers (Regular 35.3%, Micro 39.2%): roughly a third of raw, unresampled samples land legal regardless of board size or vocabulary. One reading: legality-learning difficulty is fairly stable across this project's board-size range rather than scaling sharply with board complexity.

Limitations

• Not evaluated for playing strength, deliberately.
• Smallest corpus of the three tiers (2,101 games vs. Micro's 4,135 and Regular's 15,000) — self-play on this larger, slower board took meaningfully longer per game to generate within this build's time budget.
• Legality is learned, not guaranteed — same resample-then-force-random fallback as every tier in this series.
• No weights in the tree (ADR-0002).

How to reproduce

cd projects/daydream/models/daydream-chess-nanogpt-grand-1
python prepare.py             # -> grand/{train,val}.bin + meta.pkl
python train.py config.py     # -> ./ckpt.pt (3000 iters, val ~1.05)
python harness.py --games 30  # verification

Requires Fairy-Stockfish on PATH (brew install fairy-stockfish) — Grand needs the vendored variants.ini, unlike Micro which uses a built-in variant. See ADR-0021.

Experiment write-up: Can a chess model's illegal moves be the point?

Citation / credits

• The shared core engine (modern nanoGPT lineage — RoPE, RMSNorm, bias-free).
• Fairy-Stockfish — self-play corpus generator and legality arbiter, extending its built-in grand variant.
• Set up and trained with Claude (Claude Code).