thomas-schweich/stockfish-nodes1

Stockfish Self-Play (nodes=1)

1M games of Stockfish 17 self-play at 1 node per move. Pre-tokenized in the PAWN training format.

At nodes=1, Stockfish evaluates each position with a single NNUE forward pass (no tree search). Despite the lack of search, the NNUE evaluation head produces surprisingly strong play — far from random. Games exhibit coherent openings, reasonable piece development, and tactical awareness, though blunders are more frequent than at higher node counts. This makes the dataset a useful intermediate between fully random games and strong engine play.

Schema

Column	Type	Description
tokens	list[int16]	PAWN token IDs per ply (variable length, max 255)
game_length	uint16	Number of half-moves
result	string	Game result (1-0, 0-1, 1/2-1/2, *)

Token vocabulary: 4,278 tokens (1 PAD + 4,096 grid moves + 176 promotions + 5 outcomes). See the PAWN architecture docs for details.

Usage

from datasets import load_dataset

ds = load_dataset("thomas-schweich/stockfish-nodes1")
game = ds["train"][0]
print(game["tokens"])       # [919, 3300, 659, ...]
print(game["result"])       # "0-1"
print(game["game_length"])  # 74

Or with Polars:

import polars as pl

df = pl.scan_parquet("hf://datasets/thomas-schweich/stockfish-nodes1/data/*.parquet")
print(df.head(5).collect())

Generation

Games were generated with Stockfish 17 using the Rust UCI engine interface in the PAWN repository (engine/src/engine_gen.rs).

Parameters:

• Nodes per move: 1 (single NNUE evaluation, no search)
• Opening diversity: MultiPV=5 with softmax temperature sampling throughout the full game
• Temperature: 1.0 (1 pawn difference ≈ e-fold probability ratio)
• Max ply: 500 (games exceeding this are drawn)
• Workers: 16 parallel engines, deterministic seeds (10000–10015)
• Format: zstd-compressed Parquet

License

MIT. Stockfish self-play data — no human game data.