README.md

---
license:
  - cc-by-4.0
  - odbl
task_categories:
  - reinforcement-learning
tags:
  - chess
  - training-data
  - mcts
pretty_name: Stoofvlees Training Data
---

# Stoofvlees Training Data

Chess training data produced for Stoofvlees, winner of the 2023 World Computer
Chess Championship (Valencia) and 2024 (Santiago de Compostela).

This dataset corresponds to the training data described in:

> Pascutto, G. (2025). *Sin Dolor No Hay Gloria: On Conquering 2 World Computer Chess
> Championships in Spain*. ICGA Journal, 47(3), 136–156.
> https://doi.org/10.1177/13896911251360654

## Dataset contents

| File(s) | Size | Contents |
|---|---|---|
| `train_fullgame.zdat` (+`.idx`) | ~356 MB | Games played to completion, without early adjudication/resignation |
| `train_fullgame1.zdat` (+`.idx`) | ~3.9 GB | Games played to completion, without early adjudication/resignation |
| `train_all.zdat` (+`.idx`) | ~109 GB | Training data |
| `train_early.zdat` (+`.idx`) | ~120 GB | Training data |
| `test80-2024-{mar,apr,may,jun,jul,aug,sep}.zdat` (+`.idx`) | ~72–176 GB each | Leela Chess Zero training data, converted to this format for data-mix experiments |
| `sample/sample.zdat` (+`.idx`) | ~25 GB | ~150M positions randomly sampled from the entire set, plus validation set |

Positions are subsampled rather than exhaustive; long endgames and drawn games are
under-represented relative to their share of full games.

## File format

### `.zdat` / `.zdat.idx`

`.zdat` files hold variable-length zstd-compressed groups of fixed-size position records,
compressed against a shared dictionary embedded in the file. `.zdat.idx` is a companion
index giving byte offsets for random access.

- **`.zdat` header:** magic `SJNGDAT1` (8 bytes) + `entry_size` (u32) + `batch` (u32, records
  per group) + `level` (u32, zstd compression level) + `dict_len` (u32) + a zstd dictionary
  blob of `dict_len` bytes, followed by a sequence of zstd frames — one frame per group of
  `batch` positions, each frame compressed against the dictionary.
- **`.zdat.idx` header:** magic `SJNGIDX1` (8 bytes) + record count (u64) + `count + 1`
  absolute byte offsets (u64 each) into the corresponding `.zdat` file — offset `i` is the
  start of frame `i`, offset `count` is the end of the file.

### Per-position record (uncompressed, decoded from each `.zdat` group)

Each record is a fixed 4545-byte struct, `<368s4096B18fffB` in Python `struct` notation:

| Field | Size | Description |
|---|---|---|
| bitplanes | 368 bytes | 46 bit-packed 8×8 boards, 1 byte per row, MSB-first files a–h. Always oriented from the side-to-move's perspective (board flipped if black to move). See layout below. |
| policy | 4096 bytes | MCTS visit-count policy target, one byte per (from, to) move (`index = from_square * 64 + to_square`). `0` = illegal move; otherwise the value is `1 + round(p * 254)` for policy probability `p` (decode: `p = (value - 1) / 254`). |
| scalars | 18 × float32 (72 bytes) | Auxiliary position/game-state features. See layout below. |
| score | float32 | Search/eval score for the position. |
| outcome | float32 | Final game outcome from the position's side to move: -1 (loss), 0 (draw), +1 (win). |
| moves_to_go | uint8 | Plies remaining until game end, clamped to 255. `0` means unknown/not recorded. |

Total: 368 + 4096 + 72 + 4 + 4 + 1 = 4545 bytes.

#### Bitplane layout (planes 0–45)

| Planes | Content |
|---|---|
| 0–5 | Own pawns, knights, bishops, rooks, queens, king |
| 6–11 | Opponent pawns, knights, bishops, rooks, queens, king |
| 12–21 | Attack maps for both sides across piece types |
| 22–27 | Reserved / engine-internal |
| 28 | Open files |
| 29 | Own half-open files |
| 30 | Own passed pawns |
| 31–33 | Reserved / engine-internal |
| 34 | Opponent half-open files |
| 35 | Opponent passed pawns |
| 36–38 | Reserved / engine-internal |
| 39–43 | Reserved / engine-internal |
| 44 | En passant target square (all-zero if none) |
| 45 | Light/dark square color mask (constant, not side-relative) |

#### Scalar layout (scalars 0–17)

| Index | Content |
|---|---|
| 0 | Static evaluation score, converted from centipawns to a win-probability-style value and rescaled to [-1, 1] |
| 1 | Quiescence-search score, converted the same way |
| 2–5 | Castling rights: own kingside, own queenside, opponent kingside, opponent queenside |
| 6 | Position has occurred before this game (repetition flag) |
| 7 | Fifty-move-rule counter, normalized to [0, 1] |
| 8–17 | Reserved / engine-internal |

## License

- `train_fullgame.zdat`, `train_fullgame1.zdat`, `train_all.zdat`, `train_early.zdat`: CC-BY-4.0
- `test80-2024-*`: ODbL 1.0 / DbCL 1.0 (converted from Leela Chess Zero training data,
  provided for ease of use when experimenting with data mixing)

## Citation

```bibtex
@article{pascutto2025sindolor,
  author  = {Pascutto, Gian-Carlo},
  title   = {Sin Dolor No Hay Gloria: On Conquering 2 World Computer Chess Championships in Spain},
  journal = {ICGA Journal},
  volume  = {47},
  number  = {3},
  pages   = {136--156},
  year    = {2025},
  doi     = {10.1177/13896911251360654}
}
```