kuzheren/geometry-dash-retro-tokenized-v2

Geometry Dash Retro Levels — Tokenized v2 Source

Processed Geometry Dash level-object corpus for autoregressive level generation. It is derived from kuzheren/geometry-dash-retro-levels at revision 0a42632590ce0248616d881c3e82dd80f41b80b9.

The source dataset is distributed under the MIT license. This processed version keeps the same license and records the source revision for reproducibility.

Summary

Split	Levels	Objects
train	40,905	88,006,445
validation	2,272	6,502,863
test	2,272	6,499,997
total	45,449	101,009,305

• 2,139 exact duplicate levels removed;
• 12,542 verified near-duplicate edges and 5,843 copied_id edges kept inside one split;
• 8,997 deterministic validation/test windows;
• maximum context: 2,048 objects;
• 31 object types and 96 transform states.

Repository structure

.
├── README.md
├── LICENSE
├── manifest.json
├── split_manifest.parquet
├── derived_index.parquet
├── eval_windows.npy
├── reports/
│   ├── canonical_integrity.csv
│   ├── lsh_report.json
│   ├── near_pairs.csv
│   ├── split_report.json
│   ├── vocab_v2_candidate.json
│   └── vocab_window_audit.json
└── derived_shards/
    ├── shard_00/
    │   ├── x.npy
    │   ├── y.npy
    │   ├── obj.npy
    │   ├── tr.npy
    │   ├── offsets.npy
    │   ├── level_ids.npy
    │   ├── split.npy
    │   └── done.json
    └── ... 48 shards total

Array contract

Each shard contains complete levels concatenated into flat arrays. Level i occupies offsets[i]:offsets[i + 1].

File	dtype	Meaning
x.npy	int32	X coordinate in 5-unit grid cells
y.npy	int32	Y coordinate in 5-unit grid cells
obj.npy	uint8	index into the 31-entry idx2raw table
tr.npy	uint8	flip * 24 + rotation_bin_15deg
offsets.npy	int64	level boundaries
level_ids.npy	int64	Geometry Dash level IDs
split.npy	int8	0=train, 1=validation, 2=test

Objects are stably sorted by (x, y). Full levels are stored without pre-cut training crops. A trainer samples a window of up to 2,048 objects and normalizes coordinates by subtracting the minimum X and Y within that window.

Minimal reader

from huggingface_hub import snapshot_download
from pathlib import Path
import numpy as np

root = Path(snapshot_download(
    repo_id="kuzheren/geometry-dash-retro-tokenized-v2",
    repo_type="dataset",
))

shard = root / "derived_shards" / "shard_00"
x = np.load(shard / "x.npy", mmap_mode="r")
y = np.load(shard / "y.npy", mmap_mode="r")
obj = np.load(shard / "obj.npy", mmap_mode="r")
tr = np.load(shard / "tr.npy", mmap_mode="r")
offsets = np.load(shard / "offsets.npy", mmap_mode="r")

i = 0
start, end = map(int, offsets[i:i + 2])
level = {"x": x[start:end], "y": y[start:end], "obj": obj[start:end], "tr": tr[start:end]}

Tokenized v2 vocabulary

Field	Size	Token range
PAD/BOS/EOS	3	0–2
DX	52	3–54
DY	125	55–179
Y_ABS	298	180–477
OBJ	31	478–508
TR	96	509–604

The model vocabulary is rounded to 640 tokens; 605 are assigned. Exact tables and the raw object-ID mapping are in reports/vocab_v2_candidate.json.

Deduplication and splits

Exact duplicate levels are removed. Verified near duplicates use exact Jaccard similarity >= 0.90 over translation-invariant five-object shingles. Near and direct copied_id relations cannot cross splits.

Author ID is retained in split_manifest.parquet for auditing but is not a hard split boundary. Combining author, copy and near-duplicate edges produced a single transitive component of more than 9,000 levels, so the released split prioritizes direct content-leakage protection. Author overlap is a known limitation.

Reproducibility

• canonical manifest SHA256: 4f06082663173d0e1d629914eaa08fca6b67761d722777448f66fb39cbedc544;
• postprocess report SHA256: 1342dfff83eeb1b314ba52b3227fa25e53d75d1281bcb02aaafa6e28f87cf645;
• processing schema: gd_generator_postprocess_v1.

The processed representation is intended for research on generative modeling. No claim is made that generated levels are playable or globally coherent.