Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
21
60
# K_self K_opp DB_Type(win/draw) Compression_Type Block_Size
24 24 win ZSTD 33554432
24 24 draw ZSTD 33554432
22 24 win ZSTD 33554432
22 24 draw ZSTD 33554432
24 22 win ZSTD 33554432
24 22 draw ZSTD 33554432
20 24 win ZSTD 33554432
20 24 draw ZSTD 33554432
22 22 win ZSTD 33554432
22 22 draw ZSTD 33554432
24 20 win ZSTD 33554432
24 20 draw ZSTD 33554432
18 24 win ZSTD 33554432
18 24 draw ZSTD 33554432
20 22 win ZSTD 33554432
20 22 draw ZSTD 33554432
22 20 win ZSTD 33554432
22 20 draw ZSTD 33554432
24 18 win ZSTD 33554432
24 18 draw ZSTD 33554432
16 24 win ZSTD 33554432
16 24 draw ZSTD 33554432
18 22 win ZSTD 33554432
18 22 draw ZSTD 33554432
20 20 win ZSTD 33554432
20 20 draw ZSTD 33554432
22 18 win ZSTD 33554432
22 18 draw ZSTD 33554432
24 16 win ZSTD 33554432
24 16 draw ZSTD 33554432
14 24 win ZSTD 33554432
14 24 draw ZSTD 33554432
16 22 win ZSTD 33554432
16 22 draw ZSTD 33554432
18 20 win ZSTD 33554432
18 20 draw ZSTD 33554432
20 18 win ZSTD 33554432
20 18 draw ZSTD 33554432
22 16 win ZSTD 33554432
22 16 draw ZSTD 33554432
24 14 win ZSTD 33554432
24 14 draw ZSTD 33554432
12 24 win ZSTD 33554432
12 24 draw ZSTD 33554432
14 22 win ZSTD 33554432
14 22 draw ZSTD 33554432
16 20 win ZSTD 33554432
16 20 draw ZSTD 33554432
18 18 win ZSTD 33554432
18 18 draw ZSTD 33554432
20 16 win ZSTD 33554432
20 16 draw ZSTD 33554432
22 14 win ZSTD 33554432
22 14 draw ZSTD 33554432
24 12 win ZSTD 33554432
24 12 draw ZSTD 33554432
10 24 win ZSTD 33554432
10 24 draw ZSTD 33554432
12 22 win ZSTD 33554432
12 22 draw ZSTD 33554432
14 20 win ZSTD 33554432
14 20 draw ZSTD 33554432
16 18 win ZSTD 33554432
16 18 draw ZSTD 33554432
18 16 win ZSTD 33554432
18 16 draw ZSTD 33554432
20 14 win ZSTD 33554432
20 14 draw ZSTD 33554432
22 12 win ZSTD 33554432
22 12 draw ZSTD 33554432
24 10 win ZSTD 33554432
24 10 draw ZSTD 33554432
8 24 win ZSTD 33554432
8 24 draw ZSTD 33554432
10 22 win ZSTD 33554432
10 22 draw ZSTD 33554432
12 20 win ZSTD 33554432
12 20 draw ZSTD 33554432
14 18 win ZSTD 33554432
14 18 draw ZSTD 33554432
16 16 win ZSTD 33554432
16 16 draw ZSTD 33554432
18 14 win ZSTD 33554432
18 14 draw ZSTD 33554432
20 12 win ZSTD 33554432
20 12 draw ZSTD 33554432
22 10 win ZSTD 33554432
22 10 draw ZSTD 33554432
24 8 win ZSTD 33554432
24 8 draw ZSTD 33554432
6 24 win ZSTD 33554432
6 24 draw ZSTD 33554432
8 22 win ZSTD 33554432
8 22 draw ZSTD 33554432
10 20 win ZSTD 33554432
10 20 draw ZSTD 33554432
12 18 win ZSTD 33554432
12 18 draw ZSTD 33554432
14 16 win ZSTD 33554432
End of preview. Expand in Data Studio

Bestemshe TableBase — Strong Solution Proof

Authors: Ansar Zeinulla & Murat Manassov
Affiliation: Nazarbayev University, Kazakhstan
Code Repository: github.com/ansarzeinulla/Bestemshe
Live Interactive Explorer: huggingface.co/spaces/ansarzeinulla/bestemshe-god-algorithm

Bestemshe is a traditional Kazakh two-player mancala-style game (5 pits per player, 50 total stones). This artifact is a strong solution of the game: for every reachable, legal position the exact game-theoretic value (win / draw for the side to move) has been computed by retrograde analysis and stored, enabling perfect play via direct $O(1)$ lookup without runtime search.

Together with the solver that produced it, this tablebase constitutes a machine-verifiable proof of the game's outcome under optimal play: a forced win for the second player (Follower).


Technical Summary & Compression Mechanics

A position is described by the two Kazans (captured-stone stores) k1 (side to move) and k2 (opponent), and ten pits of 0..50 stones. Material is conserved: i=110pits[i]+k1+k2=50\sum_{i=1}^{10} \text{pits}[i] + k_1 + k_2 = 50

Captures are strictly even, so each Kazan score is an even integer ($0 \le k_1, k_2 \le 24$). Positions are stored canonically from the side-to-move perspective.

Layer Layout & Storage

The dataset is sharded into 169 layers indexed by the Kazan pair (k1, k2). Each layer consists of two Zstandard-compressed bitset files:

  • layer_<k1>_<k2>_win.bin — WIN/LOSS bitset for that layer.
  • layer_<k1>_<k2>_draw.bin — DRAW bitset for that layer.

Bitsets are indexed via a colexicographical ranking function ($I_B$) that maps multiset stone distributions directly to contiguous integer offsets. Each file contains a 4-byte header specifying the number of 4 MiB decompressed blocks, followed by block offsets.

Metric Value
Layer pairs (k1, k2) 169
Total files 338 (169 win + 169 draw)
WIN files total size 6.0 GB
DRAW files total size 2.4 GB
Grand Total Size 8.3 GB (8,962,782,421 bytes)

Remote O(1) HTTP Range Query Client (No Download Required)

You do not need to download the 8.3 GB dataset to query board evaluations. Using HTTP Range headers, the following Python script fetches only the specific 4 MiB compressed block containing the target bit over the network.

Prerequisites

pip install requests zstandard

Python Query Script

"""Query the Bestemshe endgame oracle on Hugging Face by (k1, k2, pits).

Needs only: pip install requests zstandard
"""
from __future__ import annotations

from dataclasses import dataclass
from functools import lru_cache
from typing import Iterable, Literal, NamedTuple

import requests
import zstandard

BASE_URL = ("https://huggingface.co/datasets/ansarzeinulla/"
            "bestemshe-tablebase/resolve/main/layers/compressed")

TOTAL_STONES = 50          # conserved: k1 + k2 + sum(pits)
NUM_PITS = 5               # pits per side
MAX_KAZAN = 24             # largest kazan the dataset stores
WIN_THRESHOLD = 26         # a kazan this large has already won
BYTES_PER_BLOCK = 33_554_432 // 8   # 4 MiB decompressed per block

Verdict = Literal["win", "loss", "draw", "unknown"]


@dataclass(frozen=True)
class Position:
    """Side-to-move relative. pits[0:5] are the mover's own pits."""
    k1: int
    k2: int
    pits: tuple[int, ...]          # length 10

    def is_valid(self) -> bool:
        if not (0 <= self.k1 <= MAX_KAZAN and self.k1 % 2 == 0):
            return False
        if not (0 <= self.k2 <= MAX_KAZAN and self.k2 % 2 == 0):
            return False
        if len(self.pits) != 10 or any(not 0 <= p < TOTAL_STONES for p in self.pits):
            return False
        return self.k1 + self.k2 + sum(self.pits) == TOTAL_STONES


# ---------------------------------------------------------------- indexing

@lru_cache(maxsize=None)
def n_c_r(n: int, r: int) -> int:
    """Exact binomial; 0 outside the domain, matching StateIndex.h's nCr."""
    if n < 0 or r < 0 or n < r:
        return 0
    if r == 0 or n == r:
        return 1
    r = min(r, n - r)
    result = 1
    for i in range(r):
        result = result * (n - i) // (i + 1)
    return result


def min_kazan(m: int) -> int:
    """Smallest possible side-to-move kazan given m = k1 + k2 captured."""
    return max(0, m - MAX_KAZAN)


def index_state(pos: Position) -> int:
    """The position's rank inside its (k1, k2) layer.

    I_K selects the kazan slice, I_B ranks the board composition as a
    multiset — the standard stars-and-bars combinatorial rank.
    """
    m = pos.k1 + pos.k2
    i_k = (pos.k1 - min_kazan(m)) // 2

    i_b = 0
    running = 0
    for i in range(9):                      # 9, not 10: the last pit is implied
        running += pos.pits[i]
        i_b += n_c_r(i + running, i + 1)

    remaining = TOTAL_STONES - m
    return i_k * n_c_r(remaining + 9, 9) + i_b


class BitLocation(NamedTuple):
    block_index: int
    local_byte: int      # byte offset inside the DECOMPRESSED block
    bit_offset: int      # 0..7, little-endian within the byte


def bit_location(rank: int) -> BitLocation:
    byte_index = rank // 8
    return BitLocation(
        block_index=byte_index // BYTES_PER_BLOCK,
        local_byte=byte_index % BYTES_PER_BLOCK,
        bit_offset=rank % 8,
    )


def layer_urls(pos: Position, base_url: str = BASE_URL) -> tuple[str, str]:
    """(win_url, draw_url) for this position's layer."""
    return (f"{base_url}/layer_{pos.k1}_{pos.k2}_win.bin",
            f"{base_url}/layer_{pos.k1}_{pos.k2}_draw.bin")


# ------------------------------------------------------------------ moves

class Child(NamedTuple):
    pit: int                    # 0..4, the mover's own pit
    position: Position | None   # None when the move ends the game outright
    wins_immediately: bool      # opponent row emptied, or kazan >= 26


def legal_pits(pos: Position) -> list[int]:
    return [i for i in range(NUM_PITS) if pos.pits[i] > 0]


def apply_move(pos: Position, pit: int) -> Child:
    """Sow one pit and return the child ALREADY FLIPPED to the opponent's view."""
    pits = list(pos.pits)
    stones = pits[pit]
    if stones == 0:
        raise ValueError(f"pit {pit} is empty")

    pits[pit] = 0
    if stones == 1:
        last = (pit + 1) % 10
        pits[last] += 1
    else:
        pits[pit] = 1
        last = pit
        for _ in range(stones - 1):
            last = (last + 1) % 10
            pits[last] += 1

    captured = 0
    if last >= NUM_PITS and pits[last] % 2 == 0:
        captured, pits[last] = pits[last], 0

    mover_kazan = pos.k1 + captured
    opponent_starved = all(p == 0 for p in pits[NUM_PITS:])
    if opponent_starved or mover_kazan >= WIN_THRESHOLD:
        return Child(pit=pit, position=None, wins_immediately=True)

    child = Position(
        k1=pos.k2,
        k2=mover_kazan,
        pits=tuple(pits[NUM_PITS:] + pits[:NUM_PITS]),
    )
    return Child(pit=pit, position=child, wins_immediately=False)


# ----------------------------------------------------------------- oracle

INVERSE: dict[Verdict, Verdict] = {
    "win": "loss", "loss": "win", "draw": "draw", "unknown": "unknown",
}


class Oracle:
    """Reads single bits out of the remote .bin layers via HTTP Range."""

    def __init__(self, base_url: str = BASE_URL, session: requests.Session | None = None):
        self.base_url = base_url
        self.session = session or requests.Session()
        self._headers: dict[str, list[int] | None] = {}
        self._blocks: dict[tuple[str, int], bytes] = {}

    def _range_get(self, url: str, start: int, end: int) -> bytes | None:
        resp = self.session.get(url, headers={"Range": f"bytes={start}-{end}"})
        if resp.status_code == 404:
            return None
        resp.raise_for_status()
        return resp.content

    def offsets(self, url: str) -> list[int] | None:
        if url in self._headers:
            return self._headers[url]

        head = self._range_get(url, 0, 4095)
        if head is None:
            self._headers[url] = None
            return None

        num_blocks = int.from_bytes(head[:4], "little")
        need = 4 + 4 * (num_blocks + 1)
        if len(head) < need:
            head = self._range_get(url, 0, need - 1)

        table = [int.from_bytes(head[4 + 4 * i: 8 + 4 * i], "little")
                 for i in range(num_blocks + 1)]
        self._headers[url] = table
        return table

    def block(self, url: str, block_index: int) -> bytes:
        cached = self._blocks.get((url, block_index))
        if cached is not None:
            return cached

        table = self.offsets(url)
        if table is None or block_index + 1 >= len(table):
            return b""

        start, end = table[block_index], table[block_index + 1] - 1
        compressed = self._range_get(url, start, end) or b""
        data = zstandard.ZstdDecompressor().decompress(
            compressed, max_output_size=BYTES_PER_BLOCK)
        self._blocks[(url, block_index)] = data
        return data

    def locate(self, pos: Position) -> dict:
        """Resolve a position to its exact file + byte range without downloading."""
        rank = index_state(pos)
        loc = bit_location(rank)
        win_url, draw_url = layer_urls(pos, self.base_url)

        plan = {"rank": rank, "block_index": loc.block_index,
                "local_byte": loc.local_byte, "bit_offset": loc.bit_offset,
                "files": {}}
        for name, url in (("win", win_url), ("draw", draw_url)):
            table = self.offsets(url)
            if table is None:
                plan["files"][name] = {"url": url, "exists": False}
                continue
            plan["files"][name] = {
                "url": url,
                "exists": True,
                "byte_range": (table[loc.block_index],
                               table[loc.block_index + 1] - 1),
                "compressed_size": table[loc.block_index + 1] - table[loc.block_index],
            }
        return plan

    def value(self, pos: Position) -> Verdict:
        """Verdict from the SIDE TO MOVE's perspective."""
        if not pos.is_valid():
            return "unknown"

        rank = index_state(pos)
        loc = bit_location(rank)
        win_url, draw_url = layer_urls(pos, self.base_url)

        if self.offsets(draw_url) is None or self.offsets(win_url) is None:
            return self._derive(pos)

        def read_bit(url: str) -> int:
            data = self.block(url, loc.block_index)
            if loc.local_byte >= len(data):
                return 0
            return (data[loc.local_byte] >> loc.bit_offset) & 1

        if read_bit(draw_url):
            return "draw"
        return "win" if read_bit(win_url) else "loss"

    def children_values(self, pos: Position) -> list[dict]:
        """Evaluate all legal child moves."""
        out = []
        for pit in legal_pits(pos):
            child = apply_move(pos, pit)
            if child.wins_immediately:
                out.append({"pit": pit, "position": None,
                            "value": "loss", "value_for_mover": "win",
                            "terminal": True})
                continue
            child_value = self.value(child.position)
            out.append({"pit": pit, "position": child.position,
                        "value": child_value,
                        "value_for_mover": INVERSE[child_value],
                        "terminal": False})
        return out

    def _derive(self, pos: Position) -> Verdict:
        children = self.children_values(pos)
        if not children:
            return "loss"
        if any(c["value_for_mover"] == "win" for c in children):
            return "win"
        if any(c["value"] == "unknown" for c in children):
            return "unknown"
        return "draw" if any(c["value"] == "draw" for c in children) else "loss"


if __name__ == "__main__":
    start = Position(k1=0, k2=0, pits=(5,) * 10)
    oracle = Oracle()

    print("Position Byte Plan:", oracle.locate(start))
    print("Initial Position Verdict:", oracle.value(start))
    print("\nChild Move Analysis:")
    for child in oracle.children_values(start):
        print(f"  Pit {child['pit'] + 1}: Mover gets {child['value_for_mover']}")

Citation

@misc{zeinulla2026bestemshe,
  title={Strongly Solving Bestemshe: A 10-Gigabyte Retrograde Tablebase Proof},
  author={Zeinulla, Ansar and Manassov, Murat},
  year={2026},
  publisher={Hugging Face Datasets},
  howpublished={\url{https://huggingface.co/datasets/ansarzeinulla/bestemshe-tablebase}}
}
Downloads last month
124

Models trained or fine-tuned on ansarzeinulla/bestemshe-tablebase

Space using ansarzeinulla/bestemshe-tablebase 1