dataset_info:
features:
- name: text
dtype: string
- name: players
dtype: string
- name: num_moves
dtype: int16
- name: n_passes
dtype: int16
splits:
- name: train
num_bytes: 6197484135
num_examples: 20000000
- name: test
num_bytes: 1176312920
num_examples: 3796010
download_size: 2805767980
dataset_size: 7373797055
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Othello (synthetic)
Uniformly-random legal Othello games -- a regeneration of the synthetic training data from Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task (Li et al., ICLR 2023), at the paper's own scale, since the original Google Drive release is no longer reachable.
Games are generated by the original Othello engine from likenneth/othello_world (vendored in this repo under third_party/othello_world/), driven by data/othello_game.py, which seeds each game by index so the dataset is reproducible and the multiprocessing workers stay independent.
The game
- Standard 8x8 Othello: four discs pre-placed in the centre, Black moves first, a move must bracket at least one contiguous line of opponent discs (in any of the 8 directions) between the new disc and one of your own, and every bracketed line flips.
- Every move here is legal but otherwise random -- uniform over the legal moves at each position. No strategy, no engine, no human games.
- Passing. A player with no legal move passes and the opponent moves again, so the colours in a game do not strictly alternate. Never infer the mover from the move index -- use the
playerscolumn. - A game ends when neither player can move, which is a full board (60 moves) or, occasionally, 59.
Notation
Cells use the upstream repo's own string form (permit_reverse): a lowercase row letter a-h followed by a column digit 1-8, e.g. d5, with flat board index row * 8 + col. Note the row comes first, which is transposed relative to standard Othello notation; it is kept so cell names, board indices and Li et al.'s own evaluation code line up without a translation layer.
Columns
text— the moves in play order, comma separated, e.g.f5, f4, c3, .... The number of moves is the game's decision-point count.players— the colour that played each move, as a string ofB/Wof the same length, e.g.BWBWWB.... A repeated letter marks a pass by the other colour.num_moves— decision points in the game (59 or 60).n_passes— how many moves were reached because the other colour had to pass.
Splits
| Split | Games | Decision points | Avg moves/game | Games with a pass | Passes |
|---|---|---|---|---|---|
train |
20,000,000 | 1,199,496,827 | 59.97 | 6,889,947 (34.4%) | 8,924,951 (0.74%) |
test |
3,796,010 | 227,670,564 | 59.98 | 1,308,791 (34.5%) | 1,696,278 (0.75%) |
Split sizes are the paper's own (20,000,000 train, 3,796,010 validation). The two splits are disjoint by construction: train draws seeds from 0 upward, test from 1,000,000,000 upward, and games are deduplicated by move sequence across both.
Usage
from datasets import load_dataset
ds = load_dataset("cfierro/othello-synthetic")
game = ds["train"][0]
moves = [m.strip() for m in game["text"].split(",")]
colors = list(game["players"]) # "B"/"W" per move
Board state, legal moves and per-position replay: data/othello_game.py (decision_points, render).
Provenance
Code: generated with
data/push_othello_dataset.py@8bc64a1c1e.⚠️ The working tree had uncommitted changes when this dataset was generated, so the linked commit may not exactly match the code that ran. Commit the generator before generating for a faithful link.
Built: 2026-08-05T18:40:22
Command:
python data/push_othello_dataset.py --out_dir /projects/nlp/data/constanzam/knowledge-ft/othello_gen_games --num_proc 64
- Resolved arguments:
{
"repo": "cfierro/othello-synthetic",
"out_dir": "/projects/nlp/data/constanzam/knowledge-ft/othello_gen_games",
"train_games": 20000000,
"test_games": 3796010,
"test_seed_offset": 1000000000,
"num_proc": 64,
"no_dedup": false,
"skip_push": false,
"push_only": false,
"dry_run": false,
"private": false
}