manoletre/chess-puzzles-1k

Chess Puzzles 1k

A balanced benchmark of 1,000 chess puzzles, sampled from the Lichess puzzle database. I built it to compare every small model I fine-tuned to play chess on the same set (write-up: manoletre.com/chess).

What's in it

• 50 puzzle themes × 20 puzzles each. Mates by pattern (back rank, smothered, Anastasia, Arabian…), tactics (skewer, clearance, interference, x-ray, intermezzo), endgames (pawn, knight, bishop, queen, queen-and-rook), special moves (castling, en passant, underpromotion) and mates in 3–5. canonical_theme is the theme a puzzle was sampled for; Themes keeps every Lichess tag.
• Every difficulty, evenly. Ratings 399–2792, about 200 puzzles in each of five Lichess-rating quintiles (difficulty_bin: ≤980, 981–1289, 1290–1611, 1612–2000, >2000).
• No leakage. No source game (source_game_id) is shared with the training puzzles I used, so a model trained on other Lichess puzzles can still be scored fairly if you split by game too.

Columns

Same as the Lichess dump, plus three: PuzzleId, FEN, Moves, Rating, RatingDeviation, Popularity, NbPlays, Themes, GameUrl, OpeningTags, canonical_theme, difficulty_bin, source_game_id.

As in Lichess, FEN is the position before the opponent's move: the first move in Moves is the opponent's, and the solver plays every second move after it.

Scoring

I count a puzzle solved only if the model plays every solver move of the solution (as on Lichess, any move that delivers checkmate also counts), stopping at the first wrong or illegal move.

Source and license

Puzzles from the Lichess puzzle database, released under CC0 by lichess.org. This selection is also CC0.