Datasetpublic

chorcat/rukh-positions-eval

chorcat/rukh-positions-eval Distinct positions sampled from rukh-games-1800, keyed by the four-field FEN and joined with Lichess/chess-position-evaluations. One row per position with the best line and every analysed line, for value heads, reward models and preference pairs without running an engine. Part of Rukh, a chess language model built from scratch as a course on generative and agentic AI. Every derived dataset ships with the exact filters and counts of its manifest.json… See the full description on the dataset page: https://huggingface.co/datasets/chorcat/rukh-positions-eval.

sourceHugging Facecc0-1.0updated 5d agoView on Hugging Face

0likes49downloads

Dataset Card

chorcat/rukh-positions-eval

Distinct positions sampled from rukh-games-1800, keyed by the four-field FEN and joined with Lichess/chess-position-evaluations. One row per position with the best line and every analysed line, for value heads, reward models and preference pairs without running an engine.

Part of Rukh, a chess language model built from scratch as a course on generative and agentic AI. Every derived dataset ships with the exact filters and counts of its manifest.json, so it can be regenerated with rukh data positions and rukh data evals.

Files

File	Bytes	SHA-256
`positions-eval.parquet`	22138204	`8aecea4739b706bb3a9596b3b6d2e57637c6f6aced0e18184818f861a537d9eb`

Rows: 488159.

Columns

Column	Description
`fen`	pieces, side to move, castling rights and en-passant square
`game_id`	game of the first occurrence
`ply`	half-move index of the first occurrence
`last_move`	UCI move that reached the position
`result`	result of that game
`phase`	`opening` (ply <= 10), `middlegame` (14+ pieces) or `endgame`
`n_seen`	occurrences in the sampled games
`best_move`	first move of the best line for the side to move
`cp`	centipawns of the best line, from White's point of view
`mate`	moves to mate of the best line (positive: White mates)
`depth`	search depth of the best line
`knodes`	kilonodes searched
`pvs`	every line as `{move, cp, mate, depth}`, best first

Provenance and filters

Source: Lichess/standard-chess-games (positions) and Lichess/chess-position-evaluations (evaluations).

json

{
  "positions": "data/positions/positions.parquet",
  "files_present": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19
  ],
  "files_fetched_now": [
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15,
    16,
    17,
    18,
    19
  ],
  "n_files": 20,
  "coverage": 0.0976,
  "best_line": "deepest; mate for the mover first; then best cp for the side to move",
  "mate_score": 10000
}

Counts:

json

{
  "positions": 5000000,
  "evaluated": 488159,
  "lines": 3424866
}

Generated on 2026-09-19 with rukh 0.0.1.

In the course

—Built in M1 · Del PGN al tensor: los labs, the lesson that runs every command behind this repository.
—Measured in the single results table: every stage of the course, the same suite, the same day.
—Bring it to the paths the configs read: uv run rukh pull rukh-positions-eval.

License

CC0 1.0. The games, puzzles and evaluations come from the Lichess open database (https://database.lichess.org, https://huggingface.co/Lichess), published under CC0; the derived files keep that license. Please credit Lichess when you use them.