Updated the text log game
#3
by 0azad0 - opened
i have cleaned and preprocessed the game log text backgammon dataset , can i publish the dataset while citing the original dataset from ArkadiumInc/ArkadiumBackgammon ?
below is the description and screenshot of the ready dataframe :
You can access the original dataset via the link
https://huggingface.co/datasets/ArkadiumInc/ArkadiumBackgammon
Dataset Description: Cleaned Backgammon Dataset for ML/DL
This dataset contains processed expert-level human match records from the Arkadium Backgammon game logs. The dataset has been clean-mapped, normalized, and preprocessed to facilitate supervised learning tasks, specifically predicting the expert's next move (policy estimation).
Dataset Dimensions
- Total Records: 875,281 state-action transitions.
- Variables: 9 columns containing raw contextual information, board configurations, legal alternatives, and the final decision chosen by the expert.
Schema & Feature Representation
| Column Name | Data Type | Description |
|---|---|---|
game_id |
object (string) |
Unique identifier representing the source match. Each raw match file (e.g., 1205794.mat) contains multiple matches (e.g., Game 1, Game 2). To uniquely identify each sub-game, they are split and suffixed: 1205794.mat0 maps to Game 1, 1205794.mat1 maps to Game 2, etc. |
turn |
int64 |
The chronological turn counter inside that specific game. |
player |
int64 |
The active player whose turn it is to move: 0 for Player 1 (P1), and 1 for Player 2 (P2). |
dice |
list[int] |
A list of two integers representing the active dice roll for the current turn (e.g., [2, 6]). |
move |
list[tuple] |
List of individual step tuples (source, destination) actually executed by the expert player (e.g., [(24, 22), (22, 16)]). |
board_before_move |
list[int] |
A 28-element numerical array representing the state of the board immediately prior to the move. Positive integers denote the active player's checker counts, while negative integers denote the opponent's checkers. Indices 24–27 trace the bar and borne-off states. |
board_after_move |
list[int] |
A 28-element numerical array representing the state of the board directly resulting from the player's final chosen moves. |
legal_moves |
list[list[int]] |
All valid, non-redundant move configurations the active player could have chosen given the current board and dice roll. Flattened as consecutive [from, to] sequences (e.g., [[24, 22, 24, 18], [24, 22, 22, 16]]). |
chosen_idx |
int64 |
The raw target integer class label representing the index within the legal_moves list that corresponds to the expert's chosen action. |
Key Preprocessing & Cleaning Pipeline Details
- Game Split Splitting Architecture: The raw structure of having multiple games grouped inside a single log file is preserved and unrolled. Individual games are indexed and parsed separately, with the index appended directly to the filename to construct the
game_id(e.g.,filename.mat{game_index}). - State Space Representation: The classic backgammon board points (1–24) and auxiliary coordinates (Bar and Off) are normalized to a consistent 28-element 1D list structure.
- Player Perspective Normalization: Player indicators are mapped directly to binary indexes (
0or1) for continuous tensor parsing. - Move Re-alignment (Multiset Matching): Logged text moves are matched against calculated legal sequences via multiset counter logic to generate the exact ground truth index (
chosen_idx), overcoming formatting and move-ordering variations. - Invalid State Filtering: Rows featuring anomalous, empty, or unresolvable action transitions have been entirely purged to ensure high training quality.
