| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| - zh |
| pretty_name: Game-Intention Benchmark |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - game |
| - intention-forecasting |
| - esports |
| - multimodal |
| - commentary-asr |
| - video-frames |
| - cross-game |
| task_categories: |
| - text-classification |
| - token-classification |
| - video-classification |
| --- |
| |
| # Game Intention |
|
|
| ## Current data layout |
|
|
| | Directory | Contents | |
| | --- | --- | |
| | `intents/` | Supplied per-source intent annotations, including CS/Valorant normalization | |
| | `transcripts/` | ASR, alignment, VAD and anonymous commentary-speaker outputs | |
| | `data_pair/<game>/` | Paired episode assets and the integrated Naraka backfill | |
| | `derived_video/fps30/` | Dense 30fps input clips and their verified media index | |
| | `manifests/` | Dataset/split and source lists; read their documented scope before mixing | |
| | `metadata/provenance/` | Source identities, alignment evidence and ingestion provenance | |
| | `review_inputs/` | Frozen quality-review inputs, retained for reproducibility | |
|
|
| Directory names describe purpose rather than collection date. Pin an HF commit |
| for reproducibility. Earlier layouts, backups and retired inputs remain available |
| through repository history; no history was rewritten by the layout cleanup. |
|
|
| The supplied annotations are not silently flattened, relabeled or translated. |
| ASR, action and reasoning fields are annotation provenance, not mandatory model |
| inputs. Dense stored video and model-side frame sampling are separate choices. |
|
|
| The obsolete ProAct 2fps preparation copies have been retired. Their dense native |
| sources are preserved on the processing server and source URLs/provenance remain |
| in `metadata/provenance/proact/`. Paired 30fps clips are retained here. Legacy |
| frozen reference assets are not substitutes for the dense-input index. |
|
|
| ## Integrity |
|
|
| Root `index.jsonl` paths are repository-root relative. LFS assets carry `sha256`; |
| ordinary Git files carry `git_blob_id` (SHA1 of `blob <bytes>\0<contents>`). |
| All entries record exact byte sizes. `verify_index.py` supports both algorithms. |
| Root `COMPLETE.json` records the SHA256 of the index itself; neither the index |
| nor COMPLETE is included in that index, avoiding circular checksums. |
| The dense-video index continues to provide SHA256 for every video clip. |
|
|
| Source VOD count, series count, map count and episode count are different units. |
| Original IDs, clock mappings and source-group identities are retained in manifests. |
|
|
| ## Archived collection sources |
|
|
| Completed collection VODs and intermediate files are not duplicated in the |
| current tree. Their pinned historical revision and per-file sizes/hashes are |
| listed in `metadata/source_archive.json` and its index. Restore only the source |
| files needed for a new operation; current intent annotations, ASR and paired |
| training clips remain in the current tree. |
|
|
| ## Apex and Overwatch annotation inputs |
|
|
| 50 Apex candidate game intervals and 50 official Overwatch map parts with verified video/ASR pairs. Read [Apex handoff](metadata/delivery/apex/HANDOFF.md) and [Overwatch handoff](metadata/delivery/overwatch/HANDOFF.md). Intent annotations are available for both games; see the annotation QC manifests and targeted review record. Videos, transcripts, source provenance and quality/integrity evidence use stable purpose-based paths. |
|
|
| ## Apex and Overwatch intention annotations |
|
|
| The 100 source annotations are accompanied by source-relative timestamps and per-game [QC manifests](manifests/intent_qc/). These manifests exclude non-gameplay and draft episodes. Accepted Apex and Overwatch windows have verified 360p/30fps video cuts and paired annotations. See [repair and QC record](metadata/quality/apex_overwatch_intent/REPAIR.json) and [observed label values](metadata/quality/apex_overwatch_intent/OBSERVED_VOCABULARY.json). Acoustic speaker IDs are not player identities. Structural QC and targeted commentary review do not constitute full semantic validation. |
|
|
| ## Apex and Overwatch data pairs |
|
|
| 3,812 verified pairs: Apex 2,385 and Overwatch 1,427. Read [Apex manifest](manifests/apex/all.jsonl) or [Overwatch manifest](manifests/overwatch/all.jsonl). Paths are relative to the repository root. Videos are 640x360 at 30fps; model-side frame sampling remains configurable. ASR, actions and reasoning are retained as annotation provenance, not automatically supplied as model inputs. Annotation/ASR timestamps remain relative to the single-game/map source; episode videos start at zero. Subtract source_video_start to convert source timestamps to episode-video timestamps. See [pair QC](metadata/quality/apex_overwatch_pairs/QUALITY_AUDIT.json). |
|
|
| ## Review video access |
|
|
| Current, directly readable review videos and explicit clock mappings are listed in [manifests/review/all.jsonl](manifests/review/all.jsonl), with per-game lists in the same directory. See [REVIEW_VIDEOS.md](REVIEW_VIDEOS.md) and the [reader](metadata/review/code/review_videos.py). Historical absolute provenance paths are not current video locators. |
|
|
