The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/harbor/harbor.py", line 171, in _split_generators
raise DataFilesNotFoundError("No task.toml or instruction.md files found")
datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Turkish draughts NNUE data
Self-play positions of Turkish draughts (Dama) with search labels, for training NNUE evaluation nets.
Game
8x8, all 64 squares. 16 men per side on rows 2-3 and 6-7. Men move one square forward or sideways, kings move like rooks. Captures are orthogonal jumps, mandatory, maximum number of pieces must be taken, captured pieces stay on the board until the chain ends. Draw by threefold repetition or 16 moves without a capture or man move. No legal move loses.
Format
16-byte magic, then fixed-size little-endian records.
TDAMANNDATA0001\0: 38-byte recordsTDAMANNDATA0003\0: 39-byte records, addsrep
| off | field | type | meaning |
|---|---|---|---|
| 0 | wm | u64 | white men bitboard |
| 8 | wk | u64 | white kings |
| 16 | bm | u64 | black men |
| 24 | bk | u64 | black kings |
| 32 | score | i16 | search score in cp, side to move |
| 34 | result | i8 | 1 win, 0 draw, -1 loss, side to move |
| 35 | stm | u8 | 0 white, 1 black |
| 36 | hm | u8 | halfmove clock |
| 37 | na | u8 | plies since last capture or man move |
| 38 | rep | u8 | v3 only: earlier occurrences of the position in the game, 0/1/2 |
Bit i is square i, a1 = 0, h8 = 63, row-major. Board is always from white's view.
import numpy as np
REC = np.dtype([("wm","<u8"),("wk","<u8"),("bm","<u8"),("bk","<u8"),
("score","<i2"),("result","i1"),("stm","u1"),("hm","u1"),("na","u1")])
REC3 = np.dtype(REC.descr + [("rep","u1")])
magic = open(path, "rb").read(16)
recs = np.memmap(path, dtype=REC3 if magic == b"TDAMANNDATA0003\x00" else REC, mode="r", offset=16)
Labeling
- Engine plays itself, alpha-beta with NNUE eval, 5,000 nodes per move (10,000 in
deepshards). - Opening: 4 to 14 random plies, then 20 plies of temperature among near-equal root moves.
humanshards start 85 percent of games from a position out of a human game (both players 1500+). - Only quiet positions (side to move has no capture) are recorded, with the search score.
- Result: adjudicated when one side holds at least 600 cp for 4 plies, or on mate. If the 16-move counter runs out at 400 cp or more, the leading side wins. Within 50 cp is a draw. Between 50 and 400 the game is discarded.
- In true draws, up to 24 shuffling positions with
naabove 31 are written as score 0, result 0. noiseshards add 120 cp deterministic eval noise for diversity.- Positions with 7 or fewer pieces can be relabeled from an endgame tablebase at conversion. gen14 is not relabeled; the game logs (
tdg/) allow it.
gen14
2.08 billion positions, 81 GB. Produced in 3.5 hours on 832 CPU cores across three machines. Shards: wide 40 percent, human 40 percent, noise 10 percent, deep 10 percent. v3/*.v3.binpack are the records, tdg/*.tdg.gz the game logs they were derived from. A net trained one epoch on this set alone gained 63 Elo over its predecessor.
Older
gen2, gen6, gen10, gen11: 38-byte records from earlier nets, kept for ablations. mixedv14: 81 GB shuffled merge of gen14 with the older sets.
License
CC BY 4.0.
- Downloads last month
- 179