TanitDataSet-C is the commercially-clean tier of TanitDataSet. Its contents are redistributable under their upstream licenses (today: 100% comma2k19, MIT). Access is request-gated so the maintainer can see who is using it and notify consumers of corrections; the gate is an access log, NOT an additional license restriction — your rights are the upstream ones.

TanitDataSet-C — the commercially-clean tier

Seed release · 90 episodes · 15.93 GB · 14 WebDataset shards

TanitDataSet-C is the commercially-clean, redistributable tier of TanitDataSet, the camera-first autonomous-driving corpus behind the TanitAD sub-300M hierarchical latent world model. Every record is owned-safe and commercial_ok: a permissive upstream license, no share-alike, no gated, non-commercial, or refuse-class source.

The tier is a per-record stamp derived structurally from a per-source license CONSTANT (SOURCE_REGISTRY), never inferred from prose, and a hard export guard refuses egress if a single row falls outside that scope.

Read the Honest limits section before planning around this dataset. This is a seed-scale release (90 episodes, one source, one road type), not a training corpus.


Episodes	90 (train 72 · val 18)
Sources	`comma2k19` (MIT) × 90 — 100 %
License classes present	`owned-safe` × 90 — no `nc-research`, no `gated-confidential`, no `refuse`
Share-alike rows	0
Shards	14 tar (11 train + 3 val), ~1.24 GB each
Total size	15.93 GB
Frame format	`uint8 [T, 9, 256, 256]` — 100 % of records
Catalog	Hive-partitioned Parquet, 90 rows

Sources & licenses

source	license	class	`commercial_ok`	`share_alike`	episodes
`comma2k19`	MIT	`owned-safe`	✅	❌	90

The C tier admits any permissive source (MIT / Apache-2.0 / CC-BY-4.0 / OpenMDW-1.1). Today exactly one of them is built — see Honest limits.

Record schema — the world-model contract

Each episode is the byte-identical contract every TanitAD adapter emits:

frames — uint8 [T, 9, 256, 256] — a 3-frame RGB stack (9 = 3×RGB), canonicalized to f_eff ≈ 266 px (the TanitAD D-016 geometry canon).
actions — f32 [T, 2] — (steer, accel), the action applied between t and t+1.
poses — f32 [T, 4] — (x, y, yaw, v) ego trajectory.
per-episode metadata — source, license_class, license_name, commercial_ok, share_alike, split, sha256 of the frame blob, build_params_hash, native intrinsics, modality flags.

Shard layout

shards/<license_class>/<source>/<split>/shard-XXXXX.tar
└── shards/owned-safe/comma2k19/train/shard-00000.tar … shard-00010.tar   (72 eps)
└── shards/owned-safe/comma2k19/val/shard-00000.tar   … shard-00002.tar   (18 eps)

Partitioning by license_class is layer 1 of the license firewall — a share-alike source would live under a segregated sharealike/ prefix and could never share a tar with non-SA data. There is none in this release.

Each tar holds three members per episode (WebDataset convention):

{episode_id}.frames.npy    uint8 [T, 9, 256, 256]   the canonical blob
{episode_id}.motion.npz    actions / poses / timestamps
{episode_id}.meta.json     the full catalog row + provenance

Loading

A shard is a plain tar — no webdataset package required.

import io, json, tarfile, hashlib, numpy as np

with tarfile.open("shards/owned-safe/comma2k19/val/shard-00000.tar") as tf:
    blobs, metas = {}, {}
    for ti in tf:
        key, _, ext = ti.name.partition(".")
        data = tf.extractfile(ti).read()
        if ext == "frames.npy": blobs[key] = data
        elif ext == "meta.json": metas[key] = json.loads(data)

    for key, meta in metas.items():
        frames = np.load(io.BytesIO(blobs[key]), allow_pickle=False)  # [T,9,256,256]
        assert hashlib.sha256(frames.tobytes()).hexdigest() == meta["sha256"]

The catalog/ Parquet index carries one row per episode (everything except the frame blob) Hive-partitioned by license_class / source / split, so you can plan a subset with a predicate before touching a single byte of video:

import pyarrow.dataset as pads
cat = pads.dataset("catalog", partitioning="hive")
rows = cat.to_table(filter=(pads.field("split") == "val")).to_pylist()

Note: no configs: auto-loader block is declared. datasets' WebDataset builder has no decoder for the .npz motion member, so an auto-config would silently drop actions and poses. Use the snippet above.

Provenance & verification

Every episode carries a sha256 of its exact frame bytes and a build_params_hash, so a consumer can verify any shard member without rebuilding it — a rotted shard fails loudly instead of training on garbage.

Verified on 2026-07-25 immediately before this release, over the actual payload bytes (not the metadata claim):

check	result
shards present	14 / 14
episodes in payload	90 (train 72 · val 18)
`sha256` re-verified over `frames.npy` bytes	90 / 90 PASS, 0 fail
catalog ↔ payload episode-id bijection	✅ exact
catalog ↔ payload `sha256` agreement	✅ exact
frame shape / dtype uniformity	90 / 90 `[T,9,256,256]` `uint8`
distinct source corpora in payload	`{comma2k19: 90}` — only
distinct license classes in payload	`{owned-safe: 90}` — only
duplicate episode ids across shards	0
train/val episode-id overlap	0
shard-path ↔ metadata split mismatches	0
gated / `refuse` / NC / share-alike rows	0 / 0 / 0 / 0

Machine-readable: LICENSE_VERIFICATION.json (both legs — the repo's own license_guard and the independent payload audit), build_report_C.json, MANIFEST.json, BUILD_MANIFEST.json, NOTICE.

Honest limits — read this before you plan around it

This is a seed release. We would rather publish the gaps than let the size of the repo imply a corpus that does not exist.

90 episodes is seed-scale, not training-scale. This is a working proof of the schema, the license firewall and the shard/catalog contract — with real records attached. It is not enough data to train a driving world model. TanitAD's own flagship trains on a different, larger, internal corpus.
One source, one road type. All 90 episodes are comma2k19: US highway, forward camera, largely free-flow. There is no urban, no intersection, no VRU-dense, no night/adverse-weather coverage in this release, and no surround camera, LiDAR, map or route annotation.
L2D contributed 0 records — no adapter exists yet. L2D (Apache-2.0) is the source that would make this tier complete across the strategic (map / speed limit / route), tactical (CAN turn-indicator) and operative (ego trajectory) layers. It is correctly registered as shippable, but the LeRobot-v3 parquet+mp4 → 9-channel-stack adapter is ~2–3 engineering days of work that has not been done. Until it lands, L2D cannot enter the corpus. Two known traps are already recorded for whoever builds it: L2D ships no camera intrinsics (a risk to the f_eff ≈ 266 canon), and its sliding-window episodes double-count ~50 % unless de-duplicated by timestamp and split on reconstructed drives rather than episodes.
PhysicalAI-AV is deliberately excluded and always will be. TanitAD's main internal training corpus is NVIDIA's gated PhysicalAI-AV. It is gated-confidential: not redistributable, firewalled, recipe-only. It is structurally unable to become a record in this lake — the ingestor raises PermissionError — and it will never appear in this dataset or in TanitDataSet-R. Nothing here is derived from it.
Waymo Open / WOD-E2E and Waymax are refused outright, not merely excluded. Their terms follow the trained weights into the model and vehicle operation, so the contamination would survive training and no tier could contain them. They are encoded as a distinct refuse license class that raises on ingest.
The split is episode-level, not route-disjoint. The cache this build read had already lost comma2k19's route ids, so train/val were split per episode. comma2k19 is one commute route re-driven, so train and val episodes can share road segments. Do not report a generalization number from this split without saying so; rebuild from the comma2k19 origin if you need a strictly route-disjoint split.
Near-duplicates are kept on purpose, and the near-dup detector over-collapses here. A two-pass perceptual dedup flagged 67 of 90 as near-duplicates of 23 exemplars. That is a detector artifact, not duplication: the 90 have distinct ids and distinct exact-frame hashes, and near-dup pairs sit at mid-keyframe L1 ≈ 0.10 vs 0.15 for random pairs — genuinely different highway scenes. A single-keyframe 8×8 aHash with transitive union-find chains homogeneous highway footage into one smear. All 90 records ship; the exemplar flag is a sampling hint, and the repeats are wanted multi-traversal signal. Control frequency by sampling weight, not by deletion.
No semantic / VLM labels in this release. The v3 goal vocabulary (VTARGET / LONMODE / HEADWAY / lead-state / scene tags) and the Chain-of-Causation traces are designed and piloted but are not in these records. Records carry frames, actions, poses and provenance only.
Anonymization is inherited, not re-applied. comma2k19 is already publicly distributed under MIT as forward-facing US-highway dashcam footage, where PII exposure is low; no additional face/plate blurring was applied by us, and nothing was added that is not already in the upstream release. If your jurisdiction requires a face/plate pass before your redistribution, run it.
Numbers here are measured on this bundle only. No TanitAD model result is quoted on this card; model facts live in the program's model registry, and quoting them from a data card is exactly the error class this program logs.

Relationship to TanitDataSet-R

R = C ∪ NC over one schema. As of this release, TanitDataSet-R contains exactly the same 90 records as C — no non-commercial source is ingested yet, so R currently adds nothing. If you want the commercial tier, this repo is the one to use.

Citation / attribution

This dataset is a re-packaging, into the TanitAD canonical world-model contract, of publicly released data. Cite the upstream source:

@article{schafer2018commute,
  title  = {A Commute in Data: The comma2k19 Dataset},
  author = {Schafer, Harald and Santana, Eder and Haden, Andrew and Biasini, Riccardo},
  journal= {arXiv preprint arXiv:1812.05752},
  year   = {2018}
}

Attribution and per-source license text ship in NOTICE.

Built by the TanitAD Phase-A lake pipeline. Provenance travels with the data.

Downloads last month: 6

Total file size:

15.9 GB

Paper for Sayood/TanitDataSet-C

A Commute in Data: The comma2k19 Dataset

Paper • 1812.05752 • Published Dec 14, 2018

You need to agree to share your contact information to access this dataset