CASTER

CASTER (Commentary Alignment and State Tracking for Esports Research) is a multimodal dataset for observation-grounded commentary generation in StarCraft: Brood War. It aligns replay-rendered observer-view clips, synchronized observer-visible structured state, and quality-controlled commentary utterances. This release contains 239 matches (205 for training, 17 for validation, and 17 for testing) and 28,734 utterance-level clip-state-commentary records. The repository includes 28,697 replay-rendered clips. A small number of trailing commentary records occur after replay termination and therefore do not have a released clip file. Original broadcast video/audio and replay source files are not redistributed. For benchmark definitions, evaluation details, and a more detailed description of the dataset construction pipeline, please refer to the companion repository: https://github.com/kimd0/CASTER.

Repository Structure

CASTER/
|- README.md
|- splits/
|  |- train.jsonl
|  |- validation.jsonl
|  `- test.jsonl
|- match_001/
|  |- clip/
|  |  |- match_001_000.mp4
|  |  |- match_001_001.mp4
|  |  `- ...
|  |- context.json
|  |- viewport.csv
|  |- state.csv
|  `- metadata.json
|- match_002/
`- ...

Files

• clip/: replay-rendered observer-view clips aligned to utterance segments.
• context.json: the main utterance-level benchmark file. Each record contains seg_index, clip_path, speech, speech_tag, time, and events.
• viewport.csv: synchronized observer viewport trace.
• state.csv: synchronized frame-indexed replay state used to derive observer-grounded events.
• metadata.json: match-level metadata and alignment information.

Metadata Notes

• Player names are anonymized as PLAYER_1 and PLAYER_2.
• game_info.map stores the standardized map name used in the release.
• game_info.map_file preserves the original raw map filename.
• source.youtube_url points to the original public YouTube source for provenance.
• alignment stores match-level synchronization metadata between replay-derived state and commentary timing.

Usage Notes

• This release is organized by match directory rather than a single flat table. Users should read files directly from the repository hierarchy.
• context.json is the core supervision file for benchmark use.
• splits/*.jsonl provide match-level split manifests.
• Some context.json records correspond to post-game commentary after replay termination; these records may not have a matching released clip.

Intended Use

• Clip-to-Observation prediction
• Clip+Observation-to-Commentary generation
• Multimodal game understanding
• Esports commentary assistance research

License

This dataset release is intended for non-commercial research use under CC BY-NC 4.0. Original source videos remain subject to the respective rights holders' and platform terms.

Reference

If you use this dataset, please cite the accompanying paper: TBD