cs.CVOct 12, 2026
SaveEsports World Model Dataset
Organizations: King Abdullah University of Science and Technology (KAUST) · Institute of Automation, Chinese Academy of Sciences · Beijing Zhongguancun Academy & Zhongguancun Institute of Artificial Intelligence
Abstract
Interactive world models must capture how multiple participants jointly change a shared environment while observing different parts of it. Competitive games make this challenge concrete: players cooperate and compete, while fog of war, team vision and camera placement determine what each can see. Modeling these interactions requires data that connect individual observations and actions to shared state and events. We introduce the Esports World Model Dataset (EWM), a collection of human gameplay spanning StarCraft II, Dota 2 and Counter-Strike 2, with 3,053 games and maps and approximately 4,880 player-view hours, including estimated hours for the ongoing CS2 extension. It includes 4,628 completed StarCraft II views, all ten perspectives for 303 Dota 2 matches, and up to ten views per round in CS2. Across these views, replay-clock alignment connects rendered video with game-native actions, structured state and events from the same match. The three games expose complementary observation conditions: private fog of war, shared team vision and first-person viewpoints. In StarCraft II, player-specific entity records further distinguish live observations from remembered snapshots and link co-observed entities across opposing views. Together with temporal alignment, these records allow researchers to distinguish legitimate information differences from disagreement about shared state. By connecting human decisions, shared dynamics and player-specific information, EWM supports research toward jointly controlled multiplayer world models and interactive environments populated by teammates and opponents learned from human play.
Figures & tables
Figure 1: EWM records every player’s view of real competitive matches. Each column shows two players of one match at the same instant, at the native resolution: both players of a StarCraft II game (Esports World Cup 2025), a Radiant and a Dire player during a Dota 2 team fight (Esports World Cup 2026), and a Terrorist and a Counter-Terrorist in a Counter-Strike 2 round (IEM Kraków 2026). Each view is paired with the player’s actions and the engine state on the replay clock of the match.
Figure 2: Thirty player views from EWM, ten each from StarCraft II, Dota 2 and Counter-Strike 2 matches.
| Dataset | Game, scale and data source | Human | Exact actions | Engine state / events | Multi- POV | Cross- genre |
| EWM (ours) | StarCraft II, Dota 2 and CS2; 3,053 games and maps and about 4,880 player-view hours from human ladder and professional replays | ✓ | ✓ | ✓ | ✓ | ✓ |
| EgoCS-400K [ 9 ] | CS/CS2; 10,000+ hours and 400K+ first-person videos from professional match replays | ✓ | ✓ | ✓ | ✓ | – |
| MIRA [ 13 ] | Rocket League; 10,000 match-hours of bot-generated 2v2 gameplay with four synchronized views | – | ✓ | ✓ | ✓ | – |
| MultiWorld [ 31 ] | It Takes Two; 100+ hours of paired human gameplay with synchronized controls and views | ✓ | – | ✓ | ✓ | – |
| Dota2-Vis [ 5 ] | Dota 2; 144 professional matches recorded from two team perspectives | ✓ | – | ❍ | ❍ | – |
| DIAMOND / CS:GO [ 23 , 1 ] | CS:GO; 95 hours of human Deathmatch gameplay, used by DIAMOND | ✓ | ✓ | ❍ | – | – |
Table 1: Comparison with gameplay datasets. Human: recorded from human players. Exact actions: actions logged by the game rather than inferred from video. Engine state / events: game state or events from the engine. Multi-POV: synchronized views of several players in one match. Cross-genre: several game genres. ✓ yes, ❍ partial, – no.
Figure 3: Where the matches come from. StarCraft II combines tournament games with high-MMR ranked games; all Dota 2 matches and the Counter-Strike 2 extended set are professional, from the events shown and their qualifiers.
| StarCraft II | Dota 2 | Counter-Strike 2 | |
| Views | 2 per game | 10 per match | up to 10 per round |
| Video | camera view without interface, 1280 720; minimap, 256 256; 22.4 fps | Player View with interface and minimap, 1280 720, 30 fps | first-person view with interface and radar, 1280 720, 16 fps |
| Actions | unit commands (ability, target, queued flag) and camera moves | unit orders of all ten players, attributed to the issuing player | button states (move, fire, jump, duck, use, reload), view angles, raw mouse movement |
| State | observed units per view (type, owner, alliance, visibility, position, health, shields, energy, orders, buffs), resources, score, upgrades, camera | complete entity state as a lossless change stream | per-tick position, velocity, view angles, health, armor, weapon, money |
| Events | unit deaths; game result | all game messages, including the combat log | kills, damage, bomb plants, round ends |
| Clock | game loop, 22.4/s | replay tick, 30/s | replay tick, 64/s |
Table 2: Modalities per game. All streams of a match share one clock.
Figure 4: Shared data-construction stages and game-specific implementations. Native clients render player views, while game-specific extraction supplies actions and records; SC2 observations come from the replaying engine. Streams are associated on each match’s replay clock and packaged as paired outputs. The CS2 core uses four-tick frame intervals; its extension uses explicit mappings. This schematic summarizes separately implemented processes, and its game images and record tables are illustrative.
Figure 5: One StarCraft II player view at the native resolution (map Winter Madness, Player 2, Terran). The camera covers a small part of the map; the minimap is released as a separate video stream.
Figure 6: One Dota 2 Player View at the native resolution (The International 2025, match 8447715678), a third of a second before the player’s hero scores a kill. Player View mode shows the player’s own camera, interface and minimap.
Figure 7: One Counter-Strike 2 POV at the native resolution (IEM Kraków 2026, Spirit vs. FURIA, Dust2, round 3), a third of a second before the player’s kill: the first-person view with the player’s HUD.
Figure 8: StarCraft II: game loop 12,057 of a Protoss-vs-Zerg game (Kung Fu Cup 2026, map Taito Citadel); both cameras are on the same standoff. A (an Archon) and B (an Immortal) are in both players’ observations, with the same tag, position, health and shield; only the owner’s record contains their attack orders. C marks two Immortals that Player 1 sees on screen but that are under fog for Player 2: they are absent from Player 2’s observation, and the same spot in Player 2’s frame is empty.
Figure 9: Dota 2: a hero kill in a team fight (Esports World Cup 2026, match 8893208737). Top: the views of the victim (slot 4) and the killer (slot 5) at the replay tick of the kill. Bottom: the released records of this shared replay tick.
Figure 10: Dota 2: all ten Player Views at the replay tick of a hero kill in a team fight (Esports World Cup 2026, match 8893208737), Radiant in the top row and Dire in the bottom row. The victim (slot 4) and the killer (slot 5) are outlined. The released records for this shared tick appear with the enlarged pair in Figure 9 .
Figure 11: Counter-Strike 2: the killer’s and the victim’s POV in frame 952 of round 14 (IEM Kraków 2026), a quarter second before tN1R kills YEKINDAR with a MAC-10 at tick 132,372. Frame 952 of every POV of the round covers ticks 132,354–132,357; the table lists both players’ state at tick 132,356. The two players face each other (yaw and , 209 units apart along X).
Figure 12: Counter-Strike 2: all ten POVs in frame 888 of round 14 (IEM Kraków 2026), T in the top row and CT in the bottom row, a quarter second before the first kill of the round: FalleN kills donk with a P2000 at tick 132,117. The killer and the victim are outlined. Frame 888 of every POV of the round covers ticks 132,098–132,101; the table lists both players’ state at tick 132,100. FalleN stands 264 units above donk and aims steeply downward (pitch ).
Figure 13: StarCraft II gallery: one game per map, labelled with the map and the player whose camera view is shown.
Figure 14: StarCraft II gallery: both players’ views at the same game loop, 55% of the way into one game on each of six maps.
Figure 15: StarCraft II over one game: both players’ views at five instants of the Esports World Cup 2025 game of Figure 1 (Zerg vs. Protoss, map Ultralove). The log rows are read from the two trajectories at the same game loops: workers and army supply, unspent minerals and gas, value of enemy units killed so far, and the actions recorded in the preceding minute.
Figure 16: Dota 2 gallery: the killer’s Player View at hero kills in twelve matches of The International 2025, labelled with the match ID.
Figure 17: Dota 2 gallery: all ten Player Views at a hero kill in a team fight of two further matches (The International 2025 and the Road to ENC 2026 Regional Qualifiers), with the killer and the victim outlined.
Figure 18: Dota 2 over one team fight (Esports World Cup 2026, match 8893208737): the killer’s and the victim’s view at four of the five hero kills, at the replay tick of each kill. The kill log lists all five kills from the chat events, with the gold bounty, the damage dealt by the killer’s team in the preceding three seconds (combat log) and the orders the killer issued in the preceding two seconds.
Figure 19: Dota 2 over one team fight (Esports World Cup 2026, match 8893208737): all ten Player Views at three replay ticks, the start of the logged window and two of the five hero kills, with the killer and the victim outlined. The kill log in Figure 18 lists all five kills from the chat events, with the gold bounty, the damage dealt by the killer’s team in the preceding three seconds (combat log) and the orders the killer issued in the preceding two seconds.
Figure 20: Counter-Strike 2 gallery: the killer’s POV a third of a second before a kill, on maps of the released series, labelled with the map and the kill.
Figure 21: Counter-Strike 2 gallery: all ten POVs one second before the first kill of a round on two further maps (IEM Kraków 2026, G2 vs. Spirit, Anubis; BLAST Open Rotterdam 2026, 9z vs. Spirit, Dust2).
Figure 22: Counter-Strike 2 over one round (IEM Kraków 2026, Spirit vs. FURIA, Nuke, round 14): the killer’s POV a third of a second before six of the seven gun kills. The kill feed is read from the round’s event table; the last column is the killer’s health in the state table at the kill tick.
Figure 23: Counter-Strike 2 over one round (IEM Kraków 2026, Spirit vs. FURIA, Nuke, round 14): all ten POVs at three instants as the round thins out; grey tiles mark eliminated players: a POV is recorded only until two seconds after the player’s death ( Section 3.2 ). The kill feed in Figure 22 is read from the round’s event table; the last column is the killer’s health in the state table at the kill tick.
| StarCraft II | Dota 2 | Counter-Strike 2 | ||
| core | extended c | |||
| Matches | 2,320 indexed games a | 303 | 50 maps | 380 maps |
| Views per match | 2 | 10 | 10 per round | 10 per round |
| Timeline hours | 430.8 | 282.8 d | 34.8 | 265 |
| View-hours | 861.5 | 2,827.9 d | 138.4 | 1,050 |
| Structured data | 69.5M steps | 8.4B state changes | 75.0M rows | 570M rows |
Table 3: Scale of the dataset. Timeline hours count each match once; view-hours sum over player views. Structured-data units differ between games.
Figure 28: StarCraft II release metadata for 2,320 indexed games and 4,628 completed player views: game length, matchup and player MMR.
| Source | Units | Series | Teams | Players |
| StarCraft II (2,320 games) | ||||
| Ranked ladder (median MMR 5,464) | 2,076 | |||
| Rongyi Cup Season 3 (2026) | 90 | |||
| Kung Fu Cup 2026 | 51 | |||
| Esports World Cup 2025 | 49 | |||
| Source not labelled | 54 | |||
Table 4: Sources by event. Units are games (StarCraft II), matches (Dota 2) or maps (Counter-Strike 2); a series is one best-of match between two teams. Teams and players are distinct teams and player accounts.
| Check | Scope | Reported observation | Application |
| SC2 camera-command timing | 5.6M camera moves | Recorded position matches the command target at the recorded step; shuffled commands give an error of 53.4 map units. | Internal command/observation timing. |
| SC2 co-observed units | 13 games; 5.2M visible-enemy records | Engine-tag joins compare type, position, health and shield. | Comparison of shared entity fields. |
| Dota frame indexing | Six audited matches | Largest replay-time difference at a shared frame index reaches 0.83 s. | Equal frame indices are insufficient for cross-view joins. |
| Dota state reconstruction | Audited change streams | Replaying the deltas reproduces the compared entity-state fields. | Reconstruction of represented entity-state fields. |
| CS2 core alignment | Per-round frame/tick convention | One video frame corresponds to four replay ticks; views share round starts but can end at different times. | Construction of round-aligned cross-player samples. |
Table 5: Data audits and alignment checks. Scope, recorded observations and application to cross-view sampling.
More in this topic
Interactive World Models
Interactive world models predict and generate how environments change in response to actions, supporting simulation, planning, and control.
See all 122 papers