thomas-schweich/stockfish-nodes1

---
license: mit
task_categories:
- other
tags:
- chess
- stockfish
- self-play
- pawn
size_categories:
- 1M<n<10M
configs:
- config_name: default
data_files:
- split: train
path: data/train-00000-of-00001.parquet
- split: validation
path: data/validation-00000-of-00001.parquet
- split: test
path: data/test-00000-of-00001.parquet
---

# Stockfish Self-Play (nodes=1)

1M games of Stockfish 17 self-play at 1 node per move. Pre-tokenized in the [PAWN](https://github.com/thomas-schweich/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](https://github.com/thomas-schweich/PAWN/blob/main/docs/ARCHITECTURE.md) for details.

## Usage

```python
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:

```python
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](https://stockfishchess.org/) using the Rust UCI engine interface in the [PAWN](https://github.com/thomas-schweich/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.