burmich/ChessFENS

---
size_categories:
- 100M<n<1B
---
# 📘 Chess FEN Dataset with WDL & Policy Vectors

This dataset contains 732 million rows of chess positions stored in Parquet format.

Each row includes the following fields:

fen: A standard FEN string representing the current board position.

wdl: A 3-element vector [win_prob, draw_prob, lose_prob] representing the win/draw/loss probabilities for the side to move.

policy: A vector of 1858 elements representing the probability distribution over legal moves for the current position, based on the Leela Chess Zero move encoding (see below). Illegal moves are marked with -1.

game_winner: Indicates the winner of the full game ([1,0,0] for example)

## Please like if this dataset is useful to you
A like goes a long way !

## 📐 Policy Vector Format
The policy vector is aligned with the Leela-style move encoding, a fixed list of 1858 possible moves like "a1b1", "e2e4", "g1f3", etc. Each index corresponds to a unique move string from this predefined list. You can map probabilities back to move strings using the same list.

## Notes
- Feel free to download only a few files for smaller datasets
- Each file contains 1M rows
- Feel free to commit a PR with condensed and fewer parquet files
- In policy_index.py, you can find the mapping from each move index to the UCI move notation.
- Each position is viewed from the white perspective. Due to Chess symmetries, these positions can be mirrored to be black-to-move. In that case, when mirroring fen, don't forget to also mirror wdl,policy and game_winner.