Request Access to CS2Radar

CS2Radar contains round-level CS2 radar videos generated from professional match demos, together with map-level and round-level annotations. Access is subject to manual approval. Please describe your affiliation, research objectives, and intended use.

Log in or Sign Up to review the conditions and access this dataset content.

CS2Radar: A Round-Level Counter-Strike 2 Radar Video Dataset

Dataset Overview

CS2Radar is an annotated radar video dataset derived from professional Counter-Strike 2 (CS2) match demos.

The dataset provides round-level MP4 radar videos, organized by tournament match and game map, together with structured JSON annotations specifying the winning side and frame count of each recorded round.

All radar videos are independently generated from the CS2 demo (.dem) files available in the companion dataset:

PriceWang/cs2dem

The collection is designed to cover professional CS2 competitions classified as S-Tier tournaments by Liquipedia:

Liquipedia Counter-Strike S-Tier Tournaments

Unlike datasets based on first-person gameplay recordings or tournament broadcast footage, CS2Radar provides a top-down radar representation of individual rounds.

Its match-level organization and explicit round-outcome annotations support reproducible research in video understanding, tactical analysis, spatiotemporal modeling, and esports analytics.

Key Characteristics

Property Description
Game Counter-Strike 2
Modality Video and structured JSON
Video format MP4
Video perspective Spectator radar / top-down view
Sample granularity One gameplay round per video
Data source CS2 demo (.dem) files
Organization Tournament match → Map → Round
Annotation format JSON
Annotation granularity Map-level and round-level
Round annotations Winning side, frame count, video path
Winner values T, CT
Access Gated, manual approval

Data Collection and Provenance

Source Dataset

CS2Radar is derived from the following companion dataset:

PriceWang/cs2dem

The source dataset contains professional CS2 match demo files.

Radar videos in CS2Radar are generated by processing these demos and exporting the resulting radar visualizations as MP4 files.

The videos are produced independently from replay data rather than extracted from livestreams, broadcast footage, or third-party video recordings.

Tournament Coverage

The dataset is intended to cover professional CS2 matches from tournaments listed under Liquipedia's S-Tier tournament classification.

Source reference:

https://liquipedia.net/counterstrike/S-Tier_Tournaments

The repository organizes data into tournament-match directories.

Examples of tournament series represented in the repository include:

  • BetBoom Dacha Belgrade 2024
  • BetBoom Dacha Belgrade Season 2
  • BLAST Bounty 2025 Season 1 Finals
  • BLAST Bounty 2025 Season 2 Finals
  • BLAST Bounty 2026 Season 1 Finals
  • BLAST Bounty 2026 Season 2 Finals

These are examples rather than an exhaustive tournament inventory. The actual coverage of any release is defined by the files present in that repository revision.

Processing Workflow

The dataset construction process consists of:

  1. Collecting professional CS2 demo files.
  2. Processing the demo data to generate radar-view visualizations.
  3. Separating gameplay into individual rounds.
  4. Encoding each round as an MP4 video.
  5. Organizing videos by tournament match and map.
  6. Generating JSON annotations containing map identifiers, round winners, frame counts, and video paths.
  7. Uploading the videos and annotations to the dataset repository.

Dataset Organization

The repository uses a hierarchical directory structure.

Each top-level tournament-match directory contains:

  • One or more map-specific subdirectories.
  • Round-level MP4 videos within each map directory.
  • An annotation.json file describing the maps and rounds.

Real Repository Example

The following structure is based on files in the repository:

cs2radar/
└── betboom-dacha-belgrade-2024-aurora-vs-mibr-bo3-WTc3uEVsZDra2_gdtosbfG/
    ├── annotation.json
    │
    ├── aurora-vs-mibr-m1-anubis/
    │   ├── round_001.mp4
    │   ├── round_002.mp4
    │   ├── round_003.mp4
    │   ├── ...
    │   └── round_018.mp4
    │
    └── aurora-vs-mibr-m2-nuke/
        ├── round_001.mp4
        ├── round_002.mp4
        ├── round_003.mp4
        ├── ...
        └── round_023.mp4

In this example:

  • The top-level directory identifies the tournament match.
  • m1 and m2 identify the maps within the match.
  • anubis and nuke identify the played maps.
  • Each round_XXX.mp4 represents one recorded round.
  • annotation.json provides annotations for both maps.

The example contains 18 annotated rounds on Anubis and 23 annotated rounds on Nuke.

Match Directory Naming

Tournament-match directories commonly follow a structure similar to:

<tournament>-<team1>-vs-<team2>-<series-format>-<identifier>/

The directory name encodes tournament and matchup information.

The trailing identifier distinguishes source match records and should be treated as part of the directory name.

Map Directory Naming

Map directories commonly follow this convention:

<team1>-vs-<team2>-m<map-index>-<map-name>/

For example:

aurora-vs-mibr-m1-anubis/

This represents the first map of the Aurora vs MIBR series, played on Anubis.

Round Video Naming

Individual video files use zero-padded round identifiers:

round_001.mp4
round_002.mp4
round_003.mp4
...

Round numbers are interpreted within their corresponding map directory.

The full path, rather than the round filename alone, should be used to identify a unique video.

Annotation Format

Each tournament-match directory contains an annotation.json file.

The file stores a JSON array. Each array element represents a map and contains a list of annotated rounds.

Annotation Example

The following is an excerpt from an actual repository annotation file:

[
  {
    "map": "de_anubis",
    "rounds": [
      {
        "winner": "T",
        "frames": 2744,
        "file": "aurora-vs-mibr-m1-anubis/round_001.mp4"
      },
      {
        "winner": "T",
        "frames": 2596,
        "file": "aurora-vs-mibr-m1-anubis/round_002.mp4"
      },
      {
        "winner": "T",
        "frames": 2946,
        "file": "aurora-vs-mibr-m1-anubis/round_003.mp4"
      }
    ]
  }
]

This excerpt illustrates the schema; it is not the complete annotation file.

Annotation Schema

Map-Level Fields

Field Type Description
map string CS2 map identifier, such as de_anubis or de_nuke.
rounds array List of annotated rounds associated with the map.

Round-Level Fields

Field Type Description
winner string Winning side of the round: T or CT.
frames integer Number of frames in the corresponding round video.
file string Relative path to the round's MP4 video, from the tournament-match directory.

Winning Side Labels

The winner field indicates which side won the round:

Value Meaning
T Terrorists
CT Counter-Terrorists

These labels identify the winning side, not necessarily the winning organization.

Since teams may switch sides during a match, researchers should not interpret T or CT as a fixed team identity.

Frame Counts

The frames field records the number of frames associated with a round video.

For example:

{
  "winner": "T",
  "frames": 2744,
  "file": "aurora-vs-mibr-m1-anubis/round_001.mp4"
}

This entry describes a round won by the Terrorist side, with 2,744 video frames.

Video duration can be calculated from the frame count and the actual video frame rate:

duration_seconds = frames / fps

The frame rate should be read from the MP4 metadata rather than assumed from the JSON annotation.

Video-to-Annotation Mapping

Each file entry refers to an MP4 file relative to the directory containing the annotation.json file.

For example:

Match directory:
betboom-dacha-belgrade-2024-aurora-vs-mibr-bo3-WTc3uEVsZDra2_gdtosbfG/

Annotation:
annotation.json

Video:
aurora-vs-mibr-m1-anubis/round_001.mp4

Together, the match directory and relative video path uniquely identify the annotated sample.

Hugging Face Dataset Viewer

Hugging Face may automatically interpret the directory structure as a video-classification dataset.

In the automatically generated Dataset Viewer, records may appear with the fields:

Field Description
video A video from the repository.
label An automatically inferred class corresponding to the containing directory.

For example, a displayed label may resemble:

vitality-vs-liquid-m1-train

Important: The automatically inferred label is not the round-winner annotation.

The authoritative round-level annotations are stored in the repository's annotation.json files.

For supervised round-outcome analysis, researchers should join videos with their corresponding JSON entries and use winner as the target label.

The automatically generated Dataset Viewer representation should not be confused with the complete raw-file annotation schema.

Dataset Statistics

The dataset is distributed as a collection of MP4 videos and JSON annotation files.

At the time of documentation, the Hugging Face repository reports approximately 121 GB of stored data.

The automatically generated Dataset Viewer reports 2,243 rows in its available train split.

These Viewer statistics describe Hugging Face's automatically indexed representation and should not be assumed to constitute a complete count of all repository MP4 files.

For exact counts of tournament matches, maps, rounds, and annotated videos, the repository file tree and annotation.json files should be enumerated for the selected revision.

Loading the Dataset

Downloading the Repository

After receiving access approval, users can authenticate with Hugging Face:

hf auth login

Download the repository:

hf download PriceWang/cs2radar \
  --repo-type dataset \
  --local-dir ./cs2radar

Reading the Annotations

The following example reads the JSON annotations for a match and resolves each video path:

import json
from pathlib import Path

match_dir = Path(
    "cs2radar/"
    "betboom-dacha-belgrade-2024-aurora-vs-mibr-bo3-"
    "WTc3uEVsZDra2_gdtosbfG"
)

with open(match_dir / "annotation.json", "r") as f:
    annotations = json.load(f)

for map_entry in annotations:
    map_name = map_entry["map"]

    for round_entry in map_entry["rounds"]:
        video_path = match_dir / round_entry["file"]

        print({
            "map": map_name,
            "video": str(video_path),
            "winner": round_entry["winner"],
            "frames": round_entry["frames"],
        })

Building a Round-Level Annotation Index

Researchers may construct a tabular index across all match directories:

import json
from pathlib import Path

import pandas as pd

root = Path("./cs2radar")
records = []

for annotation_path in root.rglob("annotation.json"):
    match_dir = annotation_path.parent

    with open(annotation_path, "r") as f:
        maps = json.load(f)

    for map_entry in maps:
        for round_entry in map_entry["rounds"]:
            video_path = match_dir / round_entry["file"]

            records.append({
                "match": match_dir.name,
                "map": map_entry["map"],
                "video_path": str(video_path),
                "winner": round_entry["winner"],
                "frames": round_entry["frames"],
            })

df = pd.DataFrame(records)
print(df.head())
print(f"Annotated rounds: {len(df)}")

This produces an index containing:

match | map | video_path | winner | frames

The resulting index can be used for dataset exploration, preprocessing, and model training.

Research Applications

Round Outcome Classification

The winner annotation provides a binary classification target distinguishing Terrorist-side and Counter-Terrorist-side victories.

Researchers can use the complete radar video to classify the recorded outcome.

For predictive research, an appropriate observation cutoff must be introduced to prevent target leakage from the ending of the round.

Early Round Outcome Prediction

A radar video can be truncated to a prefix to investigate whether early gameplay patterns are predictive of the eventual round winner.

The existing JSON provides the final winning side as the outcome label.

Tactical Video Understanding

The top-down radar perspective can support analysis of:

  • Team positioning and movement patterns.
  • Map control and rotations.
  • Coordinated attacks and defensive responses.
  • Spatial relationships between opposing sides.
  • Temporal developments across individual rounds.

These are potential research applications. The current JSON schema does not directly supply event-level tactical labels or player trajectories.

Spatiotemporal Representation Learning

Round-level radar videos provide a structured visual setting for learning representations of gameplay evolution over time.

Potential tasks include temporal video encoding, similarity search, clustering, and sequence representation learning.

Multi-Agent Behavior Analysis

The radar perspective supports the visual study of interactions among players and teams.

Researchers may investigate cooperative and adversarial behavior, subject to the limitations of the available video information.

Esports Analytics

The combination of tournament-match organization, map identifiers, round-level videos, and outcome labels supports comparative studies of competitive CS2 gameplay.

Recommended Evaluation Practices

For machine learning experiments, researchers should consider:

  1. Match-level splitting: Avoid putting rounds from the same match into both training and evaluation sets.
  2. Tournament-level evaluation: Consider holding out entire tournaments to assess generalization.
  3. Class balance: Report the distribution of T and CT outcomes.
  4. Map variation: Evaluate performance across different CS2 maps.
  5. Temporal leakage: Avoid using frames after the prediction cutoff for early-outcome prediction tasks.
  6. Version tracking: Record the dataset revision and the exact set of matches used.

The repository's automatically generated train split is not necessarily an independently curated benchmark split.

Limitations

Annotation Scope

The released JSON annotations describe map identifiers, round winners, frame counts, and video paths.

They should not be interpreted as containing frame-level player coordinates, individual player actions, kill events, utility events, or tactical event labels.

Such information would require additional processing or annotation.

Outcome Labels

The winner field identifies the victorious side (T or CT), not a persistent team identity.

Video Perspective

Radar videos provide a top-down representation rather than the full first-person gameplay experience.

Information not visible in the radar representation may be unavailable to downstream models.

Dataset Coverage

Although the dataset targets S-Tier professional CS2 tournaments, the exact contents of each release depend on available source demos and successfully generated videos.

Coverage should be verified against the files in the selected repository revision.

Dataset Updates

Files, annotations, and coverage may change as the dataset evolves.

Researchers should record the exact revision used to ensure reproducibility.

Access Policy

Gated Access — Manual Approval

CS2Radar uses gated access.

Researchers must submit an access request through the Hugging Face repository and receive approval from the dataset maintainer before downloading restricted files.

Access is evaluated on an individual basis.

Applicants are asked to provide:

  • Full Name
  • Institution or Organization
  • Institution Website
  • Contact Email
  • Country or Region
  • Intended Use
  • Research Topic or Project Description
  • Planned Use of the Demo Files
  • Agreement to respect applicable third-party rights
  • Acknowledgement that access does not grant redistribution rights

The available intended-use categories are:

  • Academic Research
  • Education
  • Non-commercial Game Analysis
  • Commercial Research
  • Other

Approval is not guaranteed.

Licensing and Third-Party Rights

The repository currently declares the Apache License 2.0.

Any license issued by the dataset maintainer applies only to rights the maintainer is authorized to license.

The radar videos are independently generated from CS2 replay data; however, Counter-Strike 2, the underlying game content, match demos, tournament identifiers, and related materials may remain subject to third-party intellectual property rights and terms of use.

Gated access approval does not itself grant any separate rights to underlying third-party materials or authorize redistribution of content for which separate permission is required.

Users are responsible for ensuring that their use complies with applicable licenses, rights, and legal requirements.

Ethical Considerations

CS2Radar is intended to support legitimate research, educational use, and esports analysis.

Users should respect applicable third-party rights, represent the dataset and its annotations accurately, and avoid unauthorized distribution of restricted materials.

Research findings should account for the limits of radar-based observations and the distinction between round-winning sides and individual teams.

Related Resources

CS2Radar Repository

https://huggingface.co/datasets/PriceWang/cs2radar

Source Demo Dataset

https://huggingface.co/datasets/PriceWang/cs2dem

Liquipedia S-Tier Tournaments

https://liquipedia.net/counterstrike/S-Tier_Tournaments

Citation

If you use CS2Radar in your research, please cite the dataset and specify the repository revision used.

@misc{pricewang_cs2radar,
  title        = {CS2Radar: A Round-Level Counter-Strike 2 Radar Video Dataset},
  author       = {{PriceWang}},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/PriceWang/cs2radar},
  note         = {Round-level radar videos and JSON annotations derived from CS2 match demos}
}
Downloads last month
320