Datasets:

ArnieRamesh
/

CounterStrike-1K

license: cc-by-nc-4.0
pretty_name: CounterStrike-1K
task_categories:
  - video-classification
  - reinforcement-learning
  - video-to-video
  - robotics
tags:
  - counter-strike-2
  - game
  - world-model
  - video-prediction
  - action-conditioned-video
  - multi-view
  - multi-agent
  - webdataset
  - audio
  - esports
size_categories:
  - 10K<n<100K
configs:
  - config_name: manifest
    data_files:
      - split: metadata
        path: manifest.parquet
  - config_name: rounds
    data_files:
      - split: metadata
        path: round_index.parquet
  - config_name: matches
    data_files:
      - split: metadata
        path: match_index.parquet
  - config_name: subsets
    data_files:
      - split: train_10h
        path: subsets/train_10h.parquet
      - split: train_50h
        path: subsets/train_50h.parquet
      - split: train_100h
        path: subsets/train_100h.parquet
      - split: train_500h
        path: subsets/train_500h.parquet
      - split: train_1000h
        path: subsets/train_1000h.parquet
      - split: train_all
        path: subsets/train_all.parquet
      - split: dust2_100h
        path: subsets/dust2_100h.parquet
      - split: full_demo_eval
        path: subsets/full_demo_eval.parquet

CounterStrike-1K

100+ POV video wall — 1,000+ hours of professional play

1,490 rendered POV-hours · 7,347 synchronized rounds · 73,470 POV clips · 7 maps · 720p + audio

Synchronized 10-POV rounds with per-frame action overlays

10 synchronized POVs with per-frame action HUD overlays

Seven active-duty maps

_{Ancient · Anubis · Dust2 · Inferno · Mirage · Nuke · Overpass}

CounterStrike-1K is the first grounded, professional-grade Counter-Strike 2 dataset with 10 synchronized first-person perspectives per round, captured from professional match demos. It is designed for video world modeling, action-conditioned video prediction, multi-view consistency research, and audio-conditioned learning.

Quickstart

Start a fresh uv project and add the loader:

mkdir cs1k-demo && cd cs1k-demo
uv init
uv add datasets "counterstrike1k @ git+https://github.com/AnirudhhRamesh/counterstrike1k"

Using pip instead

mkdir cs1k-demo && cd cs1k-demo
python -m venv .venv && source .venv/bin/activate
pip install datasets "counterstrike1k @ git+https://github.com/AnirudhhRamesh/counterstrike1k"

Stream one sample from the public shards — no download required:

from datasets import Video, load_dataset
from counterstrike1k import decode_sample

shards = load_dataset(
    "ArnieRamesh/CounterStrike-1K-360-wds", split="train", streaming=True,
).cast_column("mp4", Video(decode=False))
sample = decode_sample(next(iter(shards)))

print(sample["actions"].shape)   # (frames,) tick, delta_pitch, delta_yaw, buttons bitmask
print(sample["state"].shape)     # (frames,) pos, view, weapon, ammo, hp, money, score, …
print(len(sample["video"]))      # mp4 bytes with synchronized audio

Browse the manifest:

import pandas as pd
from huggingface_hub import hf_hub_download

manifest = pd.read_parquet(hf_hub_download(
    "ArnieRamesh/CounterStrike-1K", "manifest.parquet", repo_type="dataset",
))
mirage_train = manifest[(manifest["map_slug"] == "mirage") & (manifest["split"] == "train")]

Tiny offline preview (one match, ~2 GB):

from counterstrike1k import load_sample
for sample in load_sample():
    print(sample["metadata"]["sample_key"])
    break

Verify that decoded actions actually align with the video — overlay_frame draws a HUD with WASD/FIRE/JUMP, mouse delta, HP/armor/money, and score onto any frame:

from counterstrike1k import overlay_frame, overlay_video

overlay_frame(sample, 60)                          # PIL.Image, ready for display()
overlay_video(sample, "debug.mp4", max_frames=192) # full debug clip

The end-to-end Jupyter walkthrough is cs2_release/quickstart.ipynb in the source repo.

Repos

Repo	Contents	Size
`ArnieRamesh/CounterStrike-1K`	Manifest, round/match indices, schema, subsets, Croissant.	~700 MB
`ArnieRamesh/CounterStrike-1K-sample`	One match-map (16 rounds, 160 POV clips).	~2 GB
`ArnieRamesh/CounterStrike-1K-360-wds`	360p WebDataset shards (recommended for training).	~1.3 TB
`ArnieRamesh/CounterStrike-1K-720-wds`	720p WebDataset shards.	~1.5 TB

What's in a sample

Each WebDataset sample is one player POV across one round.

Member	Format	Description
`mp4`	H.264 + AAC	720p or 360p video at 32 FPS with synchronized stereo game audio
`actions.bin`	packed binary, 14 B/frame	`tick`, `delta_pitch`, `delta_yaw`, 12-button bitmask
`state.bin`	packed binary, 37 B/frame	view, world position, active weapon, ammo, HP, armor, money, score, helmet/defuser/bomb
`events.json`	JSON	Sparse events: round boundaries, kills (with attacker/victim/assister `pov_idx`), bomb plant/defuse/explode, blinds
`json`	JSON	Public sample metadata: ids, alignment, alive window, weapon flags, kill counts

Buttons (bit order): FORWARD, BACK, LEFT, RIGHT, JUMP, DUCK, WALK, FIRE, RIGHTCLICK, RELOAD, INSPECT, USE.

Group POVs of one synchronized round via the shared round_id = match_{12hex}__r{round:03d}. Sample keys are match_{12hex}__r{round:03d}__p{pov:02d} with pov_idx ∈ {0..9}.

Full field-level schema is in schema/.

Splits and subsets

Splits are disjoint at the match-map level — the same match never appears in two splits.

Split	POV-hours	Match-maps	Rounds	POV clips
train	1,341.7	301	6,573	65,730
val	74.5	21	383	3,830
test	74.5	20	391	3,910

Bandwidth-friendly subsets:

ten_hours = pd.read_parquet(hf_hub_download(
    "ArnieRamesh/CounterStrike-1K", "subsets/train_10h.parquet", repo_type="dataset",
))
dust2 = pd.read_parquet(hf_hub_download(
    "ArnieRamesh/CounterStrike-1K", "subsets/dust2_100h.parquet", repo_type="dataset",
))

Available: train_10h.parquet, train_50h.parquet, train_100h.parquet, train_500h.parquet, train_1000h.parquet, train_all.parquet, dust2_100h.parquet, full_demo_eval.parquet.

Maps

Ancient · Anubis · Dust2 · Inferno · Mirage · Nuke · Overpass — all 7 active-duty competitive maps, balanced by rendered POV-frame count.

Intended uses

Action-conditioned video prediction
Game world modeling
Multi-view and multi-agent consistency evaluation
Audio-conditioned prediction
State-conditioned modeling
Representation learning

Out of scope

Re-identifying players or linking players across matches
Recovering Steam IDs or online accounts
Player profiling, ranking, anti-cheat, surveillance

Public artifacts contain no Steam IDs, online account identifiers, raw HLTV identifiers, profile URLs, player names, or chat text. pov_idx is anonymous and only stable within a single match.

Citation

@dataset{counterstrike1k2026,
  title     = {CounterStrike-1K: Synchronized Multi-POV Counter-Strike 2 for World Modeling},
  author    = {Ramesh, Anirudhh},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {1.0.0},
  url       = {https://huggingface.co/datasets/ArnieRamesh/CounterStrike-1K}
}

License

CounterStrike-1K release artifacts are distributed for non-commercial research under CC BY-NC 4.0 to the extent of the authors' rights. Counter-Strike 2 and underlying game assets remain property of Valve Corporation. Raw HLTV demo files are not redistributed.