Datasets:

iapple1
/

CoCode-World-Data

Download manifests/train.jsonl, the desired train/train-XXXXX.tar shard, and the scripts in examples/. The manifest identifies every sample's shard, exact archive members, and SHA256 hashes. Place them in their original repository-relative locations.

python examples/extract_clip.py \
  --release-root . \
  --output extracted_clip

Without --sample-id, this chooses the first manifest entry. To retrieve a specific entry, pass its complete ID using --sample-id. The destination must be new. The script verifies all five members and writes their original names. It needs only Python's standard library.

Each example contains two videos, one caption, one alignment record, and one sample metadata record:

Member suffix	Meaning
`.proxy.mp4`	Original proxy555 control video; compiler 5.5.1
`.target.mp4`	Original paired GTA RGB target video
`.caption_v1.txt`	Original English caption text
`.alignment.json`	Output-frame to source-frame mapping for each view
`.json`	Portable sample identity, source, format, quality and member metadata

Both videos are 2560 × 720 with agent_000's camera on the left and agent_001's on the right. A single clip ID is one paired dual-view training example. Keep the same left/right ownership when transforming the videos.

The TAR files are uncompressed and group samples using WebDataset-style keys. The standard-library example is the reference reader; it treats .proxy.mp4 and .target.mp4 as separate member types. manifests/train.parquet, when present, is a tabular copy of the JSONL index.

Load with Hugging Face Datasets

The dataset card selects the 25 training TAR shards. For streaming access, use:

from datasets import Video, load_dataset

ds = load_dataset(
    "iapple1/CoCode-World-Data", "train", split="train", streaming=True
)
ds = ds.cast_column("proxy.mp4", Video(decode=False))
ds = ds.cast_column("target.mp4", Video(decode=False))
example = next(iter(ds))
print(example["__key__"])
print(example["caption_v1.txt"])
proxy_bytes = example["proxy.mp4"]["bytes"]
target_bytes = example["target.mp4"]["bytes"]

The explicit casts expose video bytes without requiring a decoder. A local equivalent using the released TAR files was tested with Datasets 4.4.1: first and last shards each yielded 200 examples, and a combined cross-episode stream yielded 400. Multi-dot member fields remain distinct. That version initially infers videos as binary columns, so retain the casts when video feature typing is needed. Hub viewer conversion and automatic video previews depend on the service environment and are not part of the local verification.

Caption method identity

The release preserves caption method v1: an overall summary, a left-view timeline, and a right-view timeline. The public annotation schema is named caption-v1. No caption text was rewritten during packaging.

The reference prompt is caption_v1_prompt.txt. Internal model names and annotation labels were normalized for public distribution. The original generated caption text remains unchanged. The public prompt SHA256 is:

57c1f2626720344b12d3f67a424929c574addac3425a9d1dabaa8d8edd4e0239

The captioning procedure samples RGB at 4 FPS plus the last frame, with three proxy anchors per view. Captions and automated quality labels can contain errors; this release does not claim a complete human review. Caption times such as 5.17s correspond to the preserved 124-frame clip.

Extract the metadata you need

The metadata catalog lists archive paths, sizes, checksums, and coverage. Extract map and episode archives beneath one directory, called source here. Archives preserve the original relative directory structure, including hidden .spool directories.

mkdir -p source
tar -xzf metadata/maps/scene_001.tar.gz -C source
tar -xzf metadata/episodes/episode_001.tar.gz -C source

Replace episode_001 using the selected manifest row's source.episode. The archive path is also recorded in the row's metadata references. Keep the portable metadata/source_index/global_index.sqlite3 beside the release. Full original RGB recordings and full-length proxy videos are not part of the metadata archive; their source identities remain in the scientific provenance.

Trace a clip frame to its recorded observations

For output frame i, the alignment's left[i] and right[i] entries contain source frame_id values. Those primary keys join to frames.frame_id in the global SQLite index. The joined segment_id identifies the episode, camera owner, and metadata directory.

python examples/inspect_frame.py \
  --index metadata/source_index/global_index.sqlite3 \
  --source-root source \
  --alignment extracted_clip/clip_000001.alignment.json \
  --side left --frame 0 --records \
  --output frame_record.json

--records scans the episode's raw streams. Large episodes take longer; omit it to inspect only the index, scene path, and per-frame rendering provenance. The output includes the recorded world snapshot and any directly matching client observations, including their action, camera, agent_state, and availability fields. When multiple client rows match, the script reports all of them instead of silently selecting one.

The joins preserve distinct identities:

frame_id identifies a source frame in the release index.
segment_id plus proxy_frame_index identifies a frame in its published source segment.
source_id plus rgb_decode_index identifies the decoded frame in the original recording source.
world_capture_tick maps to the world stream's sample_seq within the episode and server clock domain.
Client matching uses the recorded render frame or client sample identity and its clock domain; legacy rows may use sample_seq_estimate.

Do not join cameras by equal local frame indices. The 24 FPS clip grid may reuse a source frame, especially for older recordings at lower native FPS. Cross-view timing remains an estimate: original nearest matches allow up to 75 ms and can be many-to-one. Camera interpolation, vehicle interpolation, and occupant smoothing are explicitly recorded derived operations. The matching raw control row is an observation, not an interpolated input command.

Reuse controls and world state

Raw client rows are under <episode-relative>/.spool/agents/<agent_id>/client.jsonl. Controls include forward/right motion, look deltas, action buttons, and keyboard/mouse aggregates. Preserve pressed, released, and held separately. Honor action.availability; absent data must not be treated as a measured zero. These controls are sampled/aggregated observations, not a complete stream of operating-system input events.

World rows are under <episode-relative>/.spool/world_state.jsonl. They record observed agents, vehicles, pedestrians, objects, environment, and lifecycle information. Completeness and availability fields are part of the data. This is an observable world-state sequence, not a complete engine save or all internal AI state.

See RENDERING.md for creating a new proxy visualization from recorded trajectories. Editing controls alone does not produce new physically consistent trajectories; that requires a simulator or a learned dynamics model.