Garf commited on
Commit
0f4b29f
·
verified ·
1 Parent(s): a761df6

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +118 -0
README.md ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license:
3
+ - cc-by-4.0
4
+ - odbl
5
+ task_categories:
6
+ - reinforcement-learning
7
+ tags:
8
+ - chess
9
+ - training-data
10
+ - mcts
11
+ pretty_name: Stoofvlees Training Data
12
+ ---
13
+
14
+ # Stoofvlees Training Data
15
+
16
+ Chess training data produced for Stoofvlees, winner of the 2023 World Computer
17
+ Chess Championship (Valencia) and 2024 (Santiago de Compostela).
18
+
19
+ This dataset corresponds to the training data described in:
20
+
21
+ > Pascutto, G. (2025). *Sin Dolor No Hay Gloria: On Conquering 2 World Computer Chess
22
+ > Championships in Spain*. ICGA Journal, 47(3), 136–156.
23
+ > https://doi.org/10.1177/13896911251360654
24
+
25
+ ## Dataset contents
26
+
27
+ | File(s) | Size | Contents |
28
+ |---|---|---|
29
+ | `train_fullgame.zdat` (+`.idx`) | ~356 MB | Games played to completion, without early adjudication/resignation |
30
+ | `train_fullgame1.zdat` (+`.idx`) | ~3.9 GB | Games played to completion, without early adjudication/resignation |
31
+ | `train_all.zdat` (+`.idx`) | ~109 GB | Training data |
32
+ | `train_early.zdat` (+`.idx`) | ~120 GB | Training data |
33
+ | `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 |
34
+
35
+ Positions are subsampled rather than exhaustive; long endgames and drawn games are
36
+ under-represented relative to their share of full games.
37
+
38
+ ## File format
39
+
40
+ ### `.zdat` / `.zdat.idx`
41
+
42
+ `.zdat` files hold variable-length zstd-compressed groups of fixed-size position records,
43
+ compressed against a shared dictionary embedded in the file. `.zdat.idx` is a companion
44
+ index giving byte offsets for random access.
45
+
46
+ - **`.zdat` header:** magic `SJNGDAT1` (8 bytes) + `entry_size` (u32) + `batch` (u32, records
47
+ per group) + `level` (u32, zstd compression level) + `dict_len` (u32) + a zstd dictionary
48
+ blob of `dict_len` bytes, followed by a sequence of zstd frames — one frame per group of
49
+ `batch` positions, each frame compressed against the dictionary.
50
+ - **`.zdat.idx` header:** magic `SJNGIDX1` (8 bytes) + record count (u64) + `count + 1`
51
+ absolute byte offsets (u64 each) into the corresponding `.zdat` file — offset `i` is the
52
+ start of frame `i`, offset `count` is the end of the file.
53
+
54
+ ### Per-position record (uncompressed, decoded from each `.zdat` group)
55
+
56
+ Each record is a fixed 4545-byte struct, `<368s4096B18fffB` in Python `struct` notation:
57
+
58
+ | Field | Size | Description |
59
+ |---|---|---|
60
+ | 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. |
61
+ | 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`). |
62
+ | scalars | 18 × float32 (72 bytes) | Auxiliary position/game-state features. See layout below. |
63
+ | score | float32 | Search/eval score for the position. |
64
+ | outcome | float32 | Final game outcome from the position's side to move: -1 (loss), 0 (draw), +1 (win). |
65
+ | moves_to_go | uint8 | Plies remaining until game end, clamped to 255. `0` means unknown/not recorded. |
66
+
67
+ Total: 368 + 4096 + 72 + 4 + 4 + 1 = 4545 bytes.
68
+
69
+ #### Bitplane layout (planes 0–45)
70
+
71
+ | Planes | Content |
72
+ |---|---|
73
+ | 0–5 | Own pawns, knights, bishops, rooks, queens, king |
74
+ | 6–11 | Opponent pawns, knights, bishops, rooks, queens, king |
75
+ | 12–21 | Attack maps for both sides across piece types |
76
+ | 22–27 | Reserved / engine-internal |
77
+ | 28 | Open files |
78
+ | 29 | Own half-open files |
79
+ | 30 | Own passed pawns |
80
+ | 31–33 | Reserved / engine-internal |
81
+ | 34 | Opponent half-open files |
82
+ | 35 | Opponent passed pawns |
83
+ | 36–38 | Reserved / engine-internal |
84
+ | 39–43 | Reserved / engine-internal |
85
+ | 44 | En passant target square (all-zero if none) |
86
+ | 45 | Light/dark square color mask (constant, not side-relative) |
87
+
88
+ #### Scalar layout (scalars 0–17)
89
+
90
+ | Index | Content |
91
+ |---|---|
92
+ | 0 | Static evaluation score, converted from centipawns to a win-probability-style value and rescaled to [-1, 1] |
93
+ | 1 | Quiescence-search score, converted the same way |
94
+ | 2–5 | Castling rights: own kingside, own queenside, opponent kingside, opponent queenside |
95
+ | 6 | Position has occurred before this game (repetition flag) |
96
+ | 7 | Fifty-move-rule counter, normalized to [0, 1] |
97
+ | 8–17 | Reserved / engine-internal |
98
+
99
+ ## License
100
+
101
+ - `train_fullgame.zdat`, `train_fullgame1.zdat`, `train_all.zdat`, `train_early.zdat`: CC-BY-4.0
102
+ - `test80-2024-*`: ODbL 1.0 / DbCL 1.0 (converted from Leela Chess Zero training data,
103
+ provided for ease of use when experimenting with data mixing)
104
+
105
+ ## Citation
106
+
107
+ ```bibtex
108
+ @article{pascutto2025sindolor,
109
+ author = {Pascutto, Gian-Carlo},
110
+ title = {Sin Dolor No Hay Gloria: On Conquering 2 World Computer Chess Championships in Spain},
111
+ journal = {ICGA Journal},
112
+ volume = {47},
113
+ number = {3},
114
+ pages = {136--156},
115
+ year = {2025},
116
+ doi = {10.1177/13896911251360654}
117
+ }
118
+ ```