Esports Trajectories and Actions (ESTA) is a large dataset of parsed Counter-Strike: Global Offensive (CSGO) demos which contain rich player action and location information. Each parsed demo:
- is a JSON compressed in
.xzformat, and is roughly 1-3 MB compressed/25-75 MB decompressed - contains spatiotemporal data on player actions (damages, kills, grenade throws, bomb plants/defuses, flashes and weapon fires)
- contains frames ("game snapshots") parsed at 2 Hz, or 2 frames per in-game second
- is parsed using the awpy parser
- is from a well-known tournament according to HLTV (all participants are known to the public. These are not private demos)
To see the structure of the JSON, visit the awpy documentation. You may also join the awpy Discord for more discussion and help with working with CSGO demofiles. If you run into issues with the dataset, feel free to ask on the Discord or open an issue.
The ESTA dataset is comprised of the "Online" (data/online/) and the "LAN" (data/lan/) subsets. "Online" contains 878 demos from top Online tournaments between January 2021 and May 2022. "Lan" contains 680 demos from top LAN tournaments between July 2021 and May 2022. The list of matches for each was compiled using HLTV.org. Visit HLTV to access a comprehensive list of tournaments and matches, along with their corresponding statistics and demos.
The ESTA data is non-exhaustive, and may lack some demos from the stated time frame or tournaments. For this reason, we do not recommend using ESTA to calculate player statistics (e.g., kills, ADR, etc.) for the time frame stated above. Instead, we recommend visiting HLTV to use their aggregation and statistics tools.
| Data | Demos | Rounds | Actions | Frames | Size |
|---|---|---|---|---|---|
| Online | 878 | 23,444 | 4.7m | 4.4m | 2.2G |
| LAN | 680 | 18,338 | 3.9m | 3.5m | 1.7G |
| Total | 1,558 | 41,782 | 8.6m | 7.9m | 3.9G |
ESTA has a variety of uses. For example, we use ESTA for win probability prediction, which we provide benchmarks for at the end of this README. ESTA may also prove useful for other prediction tasks such as trajectory prediction or unsupervised strategy identification. Outside of machine learning, ESTA may be useful for visualization, completing a class project or thesis, and for practicing data science.
You can decompress a parsed demo by running
xz --decompress demo-to-open.xz
You can install awpy (requires Golang >= 1.17) through pip by running
pip install awpy
If you want to check out what a parsed demo looks like, we recommend using the GitHub web interface to locate an example .xz, rather than cloning the whole repository, which is a few GB.
Each parsed JSON contains demo metadata (such as the map, tick rate, what competition it is from, etc.), parser parameters used from awpy, the game server parameters, and a list of game rounds. In each game round, there exists round metadata (score, end reason, teams, etc.), along with six actions (kills, damages, bomb events, grenade throws, weapon fires and flashes), as well as a list of game frames, which are effectively game snapshots containing player information. We show an example of kill and frame output below.
To understand the features available in the parsed demos, we recommend visiting the awpy documentation. As frames and events have different keys, you can see example schema in the awpy documentation. If you have any questions, feel free to drop in the Discord where a community member would be more than happy to provide support. Below is an example of the schema for a "Kill" action and a "Frame" (game snapshot).
{
"tick": 9582,
"seconds": 25.71875,
"clockTime": "01:30",
"attackerSteamID": 76561198088580941,
"attackerName": "febix",
"attackerTeam": "",
"attackerSide": "T",
"attackerX": 509.1275939941406,
"attackerY": 630.7955322265625,
"attackerZ": 86.98412322998047,
"attackerViewX": 327.601318359375,
"attackerViewY": 1.9775390625,
"victimSteamID": 76561198084596669,
"victimName": "Rullaan Spotil KANDALFWOZ ;)",
"victimTeam": "",
"victimSide": "CT",
"victimX": 781.4129638671875,
"victimY": 491.81201171875,
"victimZ": 87.30707550048828,
"victimViewX": 153.6328125,
"victimViewY": 1.0711669921875,
"assisterSteamID": null,
"assisterName": null,
"assisterTeam": null,
"assisterSide": null,
"isSuicide": false,
"isTeamkill": false,
"isWallbang": false,
"penetratedObjects": 0,
"isFirstKill": true,
"isHeadshot": true,
"victimBlinded": false,
"attackerBlinded": false,
"flashThrowerSteamID": null,
"flashThrowerName": null,
"flashThrowerTeam": null,
"flashThrowerSide": null,
"noScope": false,
"thruSmoke": false,
"distance": 305.7054888578491,
"isTrade": false,
"playerTradedName": null,
"playerTradedTeam": null,
"playerTradedSteamID": null,
"weapon": "Glock-18",
"weaponClass": "Pistols"
}{
"parseKillFrame": true,
"tick": 8174,
"seconds": 3.71875,
"clockTime": "01:52",
"t": {
"side": "T",
"teamName": "",
"teamEqVal": 3550,
"alivePlayers": 5,
"totalUtility": 1,
"players": [{
"steamID": 76561198035759667,
"name": "alo0o0o0o0",
"team": "",
"side": "T",
"x": -1179.1435546875,
"y": 483.21026611328125,
"z": -55.96875,
"velocityX": 109.83319854736328,
"velocityY": 91.87308502197266,
"velocityZ": 0,
"viewX": 56.84326171875,
"viewY": 3.33984375,
"hp": 100,
"armor": 0,
"activeWeapon": "Knife",
"totalUtility": 0,
"isAlive": true,
"isBlinded": false,
"isAirborne": false,
"isDucking": false,
"isDuckingInProgress": false,
"isUnDuckingInProgress": false,
"isDefusing": false,
"isPlanting": false,
"isReloading": false,
"isInBombZone": false,
"isInBuyZone": false,
"isStanding": true,
"isScoped": false,
"isWalking": false,
"isUnknown": false,
"inventory": [{
"weaponName": "Glock-18",
"weaponClass": "Pistols",
"ammoInMagazine": 20,
"ammoInReserve": 120
}],
"spotters": [],
"equipmentValue": 200,
"equipmentValueFreezetimeEnd": 200,
"equipmentValueRoundStart": 200,
"cash": 800,
"cashSpendThisRound": 0,
"cashSpendTotal": 0,
"hasHelmet": false,
"hasDefuse": false,
"hasBomb": false,
"ping": 32,
"zoomLevel": 0
}]
},
"ct": {},
"world": [{
"objectType": "bomb",
"x": -1215.3546142578125,
"y": 470.23406982421875,
"z": -55.96875
}],
"bombPlanted": false,
"bombsite": ""
}We provide benchmarks for win probability prediction using the ESTA data. The benchmarks are available on Google Colab. In the benchmarks, we compare gradient boosted trees (LightGBM, XGBoost) with set learning methods, like Deep Sets and Set Transformers.
We use the following hyperparameters:
- Default parameters provided by the corresponding packages for LightGBM and XGBoost.
- Default parameters for the MLP provided by the sklearn package. We use 0-1 scaling for the features.
- 10 early stopping rounds for all models.
- 100 epochs and a batch size of 32 for DeepSets and Set Transformers.
- Adam optimizer with a learning rate of 0.001.
- For DeepSets, we use one linear layer as the encoder and one linear layer for the decoder.
- For the Set Transformer, we use one ISAB layer (1 attention head, 16 induced points) as the encoder. For the decoder, we use a Pooling by Multihead Attention block (1 attention head), followed by a linear layer.
The hyperparameters are also available in the Google Colab, particularly in the second to last block. To see the architecture parameters for DeepSets and Set Transformers, see the DeepSet() and SetTransformer() functions.