chorcat/rukh-pairs-onpolicy

Preference pairs built from the moves the model itself proposes. The positions are exactly those of rukh-pairs-dpo, so the only thing that differs between the two datasets is where the two moves came from -- which is what makes a DPO run on each of them a comparison. Four moves sampled per position at temperature 1.0, scored by Stockfish at a fixed depth of 10, keeping the best and the worst when they are at least 100 centipawns apart.

Of the 13 838 positions, 6 386 produced a pair. The rest is the interesting half: 3 743 (27 %) because the model proposed a single distinct legal move, and 3 709 (27 %) because its best and worst candidate were less than a pawn apart. And only 2.9 % of these pairs involve a mate, against a third of the off-policy ones: a multi-PV search finds mates the model was never going to propose, so the two datasets are two distributions of difficulty and not two sources of one thing.

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 onpolicy.

Files

File	Bytes	SHA-256
pairs.parquet	447176	52ca00b7654329c1efe6ff84d26bc4de0117df920b4aa62dc6f8cbc2afbf23a3

Rows: 6386.

Columns

Column	Description
game_id	game the position came from, for splitting without leakage
ply	half-move index of the position
prefix	moves played up to the position (UCI, space separated)
chosen	best of the sampled moves (UCI)
rejected	worst of the sampled moves (UCI)
cp_chosen	score of chosen for the side to move (mate = +/-10000)
cp_rejected	score of rejected for the side to move
phase	opening, middlegame or endgame
white_elo	Elo header the model was prompted with
black_elo	Elo header the model was prompted with
candidates	distinct legal moves the model proposed here (2 to 4)

Provenance and filters

Source: rukh-pairs-dpo for the positions, chorcat/rukh-medium for the moves.

{
  "source": "data/pairs/dpo-prompts.parquet",
  "checkpoint": "checkpoints/medium-v4-20260919-174623/best.pt",
  "candidates": 4,
  "temperature": 1.0,
  "top_k": 20,
  "engine_depth": 10,
  "min_delta_cp": 100,
  "seed": 42,
  "skipped_all_same": 3743,
  "skipped_too_close": 3709,
  "note": "the moves come from the model, the positions from the off-policy pairs; the engine is limited by depth so the dataset is rebuildable (D-107)"
}

Counts:

{
  "pairs": 6386,
  "positions": 13838
}

Generated on 2026-09-20 with rukh 0.0.1.

In the course

• Built in M5 · Alineamiento: 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-pairs-onpolicy.

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.