| tags: | |
| - trl | |
| license: cc0-1.0 | |
| # Chess Puzzles Dataset | |
| ## Summary | |
| The Chess Puzzles dataset is a processed version of the [Lichess puzzle database](https://huggingface.co/datasets/Lichess/chess-puzzles): 6 million positions from real games, each with a single winning move, rated by how hard Lichess players found them. Each puzzle is a prompt showing the board, its FEN and the legal moves, and a completion with the first move of the solution in UCI notation. | |
| ## Data Structure | |
| - **Format**: [Conversational](https://huggingface.co/docs/trl/main/dataset_formats#conversational) | |
| - **Type**: [Prompt-completion](https://huggingface.co/docs/trl/main/dataset_formats#prompt-completion) | |
| Columns: | |
| - `"prompt"`: The puzzle: the side to move, the board, its FEN and the legal moves. | |
| - `"completion"`: The first move of the solution, in UCI notation (for example `e2e4`). | |
| - `"PuzzleId"`, `"Rating"`, `"Themes"`: The Lichess puzzle ID, its rating, and its themes (`mateIn2`, `fork`, ...). | |
| The test split holds 1,000 puzzles. Lichess also accepts any move that mates, so a model that answers with another mating move solves the puzzle too. | |
| ## Generation script | |
| The script used to generate this dataset can be found [here](https://github.com/huggingface/trl/blob/main/examples/datasets/chess_puzzles.py). | |