README.md

---
license: other
license_name: mixed-cc0-1.0-and-mit
license_link: https://huggingface.co/datasets/Alfredvc/chess-autocomplete-eval-datasets/blob/main/README.md#licenses
pretty_name: Chess Autocomplete Eval Battery Datasets
tags:
  - chess
  - lichess
  - evaluation
  - benchmark
size_categories:
  - 1M<n<10M
configs:
  - config_name: maia_table1
    data_files:
      - split: test
        path: maia/example_test_dataset.csv
  - config_name: stratified_metadata
    data_files:
      - split: metadata
        path: stratified_human/merged-metadata.parquet
---

# Chess Autocomplete Eval Battery Datasets

Input data for the full evaluation suite of the chess-autocomplete releases
([91M](https://huggingface.co/Alfredvc/chess-autocomplete-v1-91m),
[350M](https://huggingface.co/Alfredvc/chess-autocomplete-v1-350m),
[700M](https://huggingface.co/Alfredvc/chess-autocomplete-v1-700m-300k)).
Inputs only: no model, eval code, or results. One directory per suite component.

```
README.md
index.json
LICENSES/
  allie-MIT.txt
  maia2-MIT.txt
allie_benchmark/
  2022-test-annotated.jsonl
maia/
  cross_skill_2023-12.jsonl
  example_test_dataset.csv
stratified_human/
  merged-map.bin
  merged-metadata.parquet
  merged.bin
```

Total size: 5.6 GiB (6.0 GB).

## stratified_human/

- Files: `merged.bin`, `merged-map.bin`, `merged-metadata.parquet`.
- Contents: 768,000 Lichess standard games from 2022 and 2023, 45,728,786 plies.
  Stratified into 768 cells: 16 White-Elo buckets by 16 Black-Elo buckets by 3
  time controls (bullet, blitz, rapid), 1,000 games per cell. Classical
  excluded.
- Encoding: same binary token format as
  [`Alfredvc/chess-autocomplete-lichess`](https://huggingface.co/datasets/Alfredvc/chess-autocomplete-lichess).
  `merged.bin` holds 46,496,786 `uint16` tokens, one per ply plus one game-end
  token per game; `merged-map.bin` holds `uint64` per-game byte offsets (768,000
  entries); `merged-metadata.parquet` holds one row per game.
- Generation: the Rust `chess-bin-stratify` tool
  (`rust/src/bin/chess_bin_stratify.rs`) re-buckets per-division binary triplets
  that `chess-pgn-stratify` encoded from the monthly Lichess `.pgn.zst` archives.
- Read by: the `stratified_human` component
  (`chess_autocomplete/evals/components/stratified_human.py`).
- License: CC0-1.0. Derived from Lichess CC0 data.

## allie_benchmark/

- File: `2022-test-annotated.jsonl`.
- Contents: 884,049 scored positions from Lichess 2022 blitz games. Opening plies
  and moves played with under 30 seconds on the clock are excluded.
- Source: file `lichess-2022-blitz-test/2022-test-annotated.jsonl` from the
  Hugging Face dataset
  [`yimingzhang/allie-data`](https://huggingface.co/datasets/yimingzhang/allie-data),
  fetched by `scripts/download_allie_test_set.py`.
- Read by: the `allie_benchmark` component
  (`chess_autocomplete/evals/components/allie/`).
- License: MIT (Allie, Zhang et al., 2024). Underlying games are Lichess CC0. See
  `LICENSES/allie-MIT.txt`.

## maia/

`example_test_dataset.csv`

- Contents: 127,852 rows of scored positions from Lichess December 2019 rapid
  games, sub-30-second moves already excluded. The `maia` component scores the
  106,740 of them with `move_ply > 10`, dropping plies 0 through 10. Grouped
  Skilled (56,812), Advanced (41,747), Master (8,181).
- Source: the Maia-2 project copy, Google Drive id
  `1fSu4Yp8uYj7xocbHAbjBP6DthsgiJy9X`.
- Read by: the `maia` component, surface `table1`.
- License: MIT (Maia-2, CSSLab, 2024). Underlying games are Lichess CC0. See
  `LICENSES/maia2-MIT.txt`.

`cross_skill_2023-12.jsonl`

- Contents: 4,932,999 scored positions from Lichess December 2023 rapid games.
  Per active-Elo by opponent-Elo cell, counts match the Maia-2 Appendix Table 8
  matrix. Row schema tag `chess_autocomplete.maia2_cross_skill_regenerated.v1`.
- Generation: `generate_maia_cross_skill_rows` in
  `scripts/run_eval_suite.py` scans
  `lichess_db_standard_rated_2023-12.pgn.zst` from
  [database.lichess.org](https://database.lichess.org/) in source order and keeps
  rows per cell until each target count is met.
- Read by: the `maia` component, surface `cross_skill_figure2`.
- License: CC0-1.0. Derived from Lichess CC0 data.

## Licenses

| Path | License | Source | Notice |
| --- | --- | --- | --- |
| `stratified_human/*` | CC0-1.0 | Lichess DB, re-encoded | |
| `maia/cross_skill_2023-12.jsonl` | CC0-1.0 | Lichess DB, regenerated | |
| `allie_benchmark/2022-test-annotated.jsonl` | MIT | Allie (Zhang et al., 2024) | `LICENSES/allie-MIT.txt` |
| `maia/example_test_dataset.csv` | MIT | Maia-2 (CSSLab, 2024) | `LICENSES/maia2-MIT.txt` |

The Allie and Maia sets are redistributed under their MIT terms. The Lichess
database is released under CC0.

## Training exclusion

The released models are not trained on the Lichess months these sets cover:
2019-12, 2023-12, and 2022. Keeping those months out of training is why this repo
exists. All three releases (91M, 350M, 700M) train on the same 21 monthly shards,
2024-01 through 2025-09, listed under `data_config.shards` in each run's
`training_config.yaml` in
[`chess-autocomplete-v1-checkpoints`](https://huggingface.co/Alfredvc/chess-autocomplete-v1-checkpoints).
Every evaluation month here predates that window.

## Download

```bash
hf download Alfredvc/chess-autocomplete-eval-datasets \
    --repo-type dataset --local-dir eval-datasets
```

The eval suite (`scripts/run_eval_suite.py`) fetches each file from this repo when
it is not present locally.

## File manifest

| file | size | sha256 |
| --- | --- | --- |
| `LICENSES/allie-MIT.txt` | 1.0 KiB | `de02e92381b5ccc5…` |
| `LICENSES/maia2-MIT.txt` | 1.0 KiB | `dad60c955a91b9d9…` |
| `allie_benchmark/2022-test-annotated.jsonl` | 42.7 MiB | `a6014de8ef861b2e…` |
| `maia/cross_skill_2023-12.jsonl` | 5.5 GiB | `17eb3c88a61e3837…` |
| `maia/example_test_dataset.csv` | 21.5 MiB | `cd4defb7213f052e…` |
| `stratified_human/merged-map.bin` | 5.9 MiB | `b47598f23f8fd0a1…` |
| `stratified_human/merged-metadata.parquet` | 487.6 KiB | `9ab335ccdcd59fc1…` |
| `stratified_human/merged.bin` | 88.7 MiB | `f0e6df96e931db4e…` |