ZiyangYe/WorldCast-data

WorldCast evaluation data

The data the evaluation of WorldCast: Distributed Multiplayer World Models reads, as the paper's own files: the latents, labels and tick tables of its three evaluation sets. With the code and the weights they run the paper's evaluations without downloading or encoding video, and the latents are the bytes the paper's runs read, so the results are bit-exact on the paper's stack.

Every file is derived from the OpenCS2 dataset by Julien Blanchon, released under CC BY 4.0, at revision 8d7deddcd30c081fb197b75f58b4a0fba1117b9d. This dataset is not affiliated with or endorsed by the author of OpenCS2, by Valve or by HLTV. Counter-Strike 2 and its maps are Valve's; no game file is included.

The sets

set	index (in the code repository)	what
validation	configs/data/validation_index.jsonl	the 64 validation windows, 16 per map (tools/evaluate.py)
windows	configs/eval/window_views.jsonl	the 512 views of the 256 ten-second windows of Tables 1 and 2
rounds	configs/eval/round_index.jsonl	the 96 clients of the 32 whole rounds of Table 3

All of them come from the 40 held-out matches (configs/data/held_out_matches.txt).

Files

Everything is under eval/; eval/manifest.json lists every file with its size and sha256, and the windows of each set.

path	files	size	what
latents/<media_id>.npz	658	2.0 GB	the windows of the latent cache the sets read: member win_<start:06d>, float16 [1, 41, 48, 24, 42]
vislabels/<media_id>.npz	658	17.6 MB	GT visibility labels of each client and view
obslabels/flashlabels/, obslabels/scopelabels/	658 + 658	7.9 MB	flash and scope labels
opencs2/rounds/.../ticks.parquet	3,280	513 MB	the tick tables of every player of every round the sets touch
media_index.jsonl	1	1.5 MB	the media index rows of those players

In five rounds of the views the paper's media index has no row for one recorded player: the paper's generation runs read that player as absent, and the entity criterion reads its tick table.

What was changed

• Latents: the OpenCS2 videos, every second source frame (16 fps), resized bilinearly to 384 x 672, encoded with the VAE of Wan2.2-TI2V-5B in bf16 and stored as float16. One window is 161 frames, 41 latent frames.
• GT visibility labels: computed by engine line-of-sight tests between the players of the recordings, against meshes exported from the game's map files. The meshes are not included.
• Flash and scope labels: computed from the recorded video and the tick tables.
• Tick tables: unchanged copies of OpenCS2's files.
• Media index: one row per recording, derived from OpenCS2's index and tick tables.

Terms

The OpenCS2 dataset card states that its .dem source data is mirrored from HLTV, that downstream use is bound by the original tournament terms, and that renders and metadata are released under CC BY 4.0. The same applies to this derived data. Please credit OpenCS2 and WorldCast when you use it.

Use

git clone https://github.com/Ziyang-Ye/WorldCast && cd WorldCast
python tools/download_eval_data.py --out-dir data/evaluation rounds    # and/or validation, windows

The tool fetches only the files of the sets it is given, checks every file against eval/manifest.json and writes data/evaluation/eval/<set>.yaml, the data paths of the set for the code's --config. The repository's docs/reproduce.md runs Table 3 with it, and docs/training.md the validation windows and the paper's metrics.

Citation

@article{ye2026worldcast,
  title   = {WorldCast: Distributed Multiplayer World Models},
  author  = {Ye, Ziyang and Huang, Junchao and Zhang, Evelyn and Xie, Zhihao and Zhang, Ruicheng and Han, Boyao and Ban, Litao and Wang, Ziye and Hu, Xinting and Shi, Shaoshuai and Tian, Zhuotao and Jiang, Li},
  journal = {arXiv preprint arXiv:2610.12412},
  year    = {2026}
}

For OpenCS2, cite as its dataset card asks.