---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: ctx
    dtype: string
  - name: target
    dtype: string
  splits:
  - name: train
    num_bytes: 748604106
    num_examples: 1479018
  download_size: 407484559
  dataset_size: 748604106
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

This is a collection of ~1.5M chess puzzles from the [Lichess database](https://database.lichess.org/#puzzles) of ~3.9M puzzles (as of 2024-05-09).

The set of puzzles from ["Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks"](https://github.com/aks2203/easy-to-hard-data/tree/main) is included, with the exception of 26,079 puzzles that are no longer in the Lichess database (on the assumption that they might have been removed for a good reason).

For each puzzle, `ctx` is a SAN transcript (with every half-move numbered) of an actual Lichess game, up to the puzzle position. Note that this includes the first move of the `Moves` column in the Lichess and Easy-to-Hard datasets.

`target` is the **best** next move, in SAN, with a leading space. This move (second move in `Moves` column) generally differs from the actual Lichess game, which may contain blunders.

Additional moves of the puzzle solution are not included.

This format matches that used in ["Weak-to-strong generalization"](https://openai.com/index/weak-to-strong-generalization/) and the set of puzzles is also intended to be as similar as possible (except for the 26k that Lichess removed).