Dataset Viewer
The dataset viewer is not available for this dataset.
Job has been terminated due to a temporary spike in resource usage and may be restarted later.
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Curated gameplay export — 2026-10-21

Target layout: michaelu0826/GTAworldmodel_v2. This export contains the previously selected Minecraft, Counter-Strike and It Takes Two subsets. Existing GTA data is not copied or changed.

Output on 23501: /home/wangyue/lumingchun/curated_gtaformat_20261021. Controller on 23701: /home/wangyue/lumingchun/curated_export_20261021. Read progress.json for current counts by dataset/status. Only an agent_NNN.complete.json marker confirms all files for that agent have passed transfer hash checks. An episode/chunk is ready only when all expected agents have completion markers.

Layout

ep_<source>_<recording_hash>/video_<source_start_milliseconds>/agent_NNN with suffixes:

  • .mp4: H264, 1280×720, 20 FPS, aspect ratio preserved with letterboxing; bitrate capped at 4 Mbps.
  • .actions.jsonl: native source actions and timestamps in raw_action; GTA 34-dimensional placeholders use false masks because cross-game semantic equivalence is not asserted. No unsupported action is represented as a known zero. X-EGO-CS rows retain trajectories rather than inventing controls. MultiWorld keyboard/gamepad ownership relative to split-screen side is unverified and explicitly marked.
  • .frames.parquet: one row per delivered frame with source frame/time, action row link and camera index.
  • .metadata.json, .previous_context.json, .video_validation.json.
  • .camera.npz: intrinsics[N,3,3] in delivered-video pixels, extrinsics[N,4,4] world-to-camera, inverse camera_to_world, time_sec, image_size_wh, and estimated-camera keyframe times.
  • .camera.json: conventions, provenance and quality status.
  • .complete.json: SHA256 and byte size for every delivered file.

Selected intervals merge only when they are contiguous within the identical recording and view set. No gaps or unrelated recordings are joined. Each chunk is at most 300 seconds; the longest contiguous selected run in this manifest is 180 seconds. Source-derived chunk identifiers preserve original time rather than pretending the recording starts at zero. Per-part duration rounds down to the 20 FPS grid; losses are recorded in video validation.

Camera semantics

Polis and fixed-skin Polis use native normalized intrinsics converted to output pixels and the original global extrinsics. Source world coordinates are retained across agents. Original per-frame parameters are resampled using the explicit selected source frame indices.

Other sources use standard FlashVGGT, camera-only BF16, 2 FPS keyframes, windows of 64 frames with 16-frame overlap. Windows are aligned using overlapping camera orientation and centers with a bounded similarity scale. Dense 20 FPS cameras use rotation SLERP and linear center/intrinsic interpolation. These are estimated_unvalidated, not measured ground truth. Monocular scale is arbitrary; each agent/clip has its own coordinate frame. Cross-agent shared-world alignment and per-frame visual refinement are not established by this pipeline. The uniform schema does not imply uniform accuracy or a common world frame.

Actions preserve native values; labels used for balancing are not presented as sensor signals. This export does not copy large voxel/instance-mask tensors. Source pointers and original selection manifests are retained for additional signal extraction.

Resource and recovery policy

12 conversion workers (2 threads each), 2 GPU workers (one per RTX4090), 4 native-camera upload workers. At most 12 tasks are in flight, with at most 4 non-native tasks waiting/running for GPU processing. Source videos and output staging use /tmp tmpfs, never the nearly full dataset disk. New conversion pauses below 8 GiB disk free, 20 GiB tmpfs free, 16 GiB available RAM, or above 24 GiB staging usage. Consumers keep draining the queue. Transfer pauses below 50 GiB free on 23501. Successful output is deleted from tmpfs only after remote SHA256 verification and completion marker creation.

Conversion/camera failures retry up to three times and remain recorded as failed rather than falsely completed. Network/storage interruptions retry without discarding completed staged results. Logs rotate. PAUSE in the controller directory pauses claiming new work; existing transfer retries continue. Authentication is supplied interactively and held only in process memory; no password is written to disk. Restarting requires interactive authentication and recovery of interrupted claims before relaunch.

Solaris source exception: 65 selected agent outputs have video and episode metadata but no action JSON in their source archives. Their .actions.jsonl files are empty, action_records=0, and source_action_status=missing_in_source_archive in metadata. These outputs retain video and camera data; no actions are invented.

Downloads last month
1,219