This dataset contains Matterport3D-derived observations. Read the applicable agreement at https://matterport.com/legal/matterport-end-user-license-agreement-academic-use-model-data and the dataset LICENSE. Your Hugging Face username, email and the information submitted below are recorded by the dataset publisher. The Matterport agreement also calls for sharing acceptance information with Matterport, subject to applicable law.

CoFL-S training dataset

This dataset supplies the prepared RGB-D observations and instruction-conditioned sector flow-field annotations used for CoFL-S. The first release covers the CoFL-S R2R/RxR collection. CoFL image-field and ScanNet datasets are outside this release.

Publication status: complete. All 92 archive parts are uploaded. Every packaged file payload was re-read and verified against the bytes read from the source; every Hub archive was checked against its size and SHA-256. See the archive manifest and checksums.

Download and extract

Accept the access terms on this page, then use the Hugging Face CLI. Choose a full commit hash from the repository history for an immutable download and pass it to --revision in every download command.

hf auth login
hf download lhk66666/CoFL-S-Dataset --repo-type dataset --local-dir cofl-s-download
(cd cofl-s-download && sha256sum -c SHA256SUMS)
mkdir -p datasets/cofl-s
for part in cofl-s-download/archives/*.tar; do
  tar -xf "$part" -C datasets/cofl-s
done

Keeping all archives and the extracted collection requires roughly 720 GB of free space. With less space, download, verify and extract one named archive at a time, then remove that downloaded archive. Always include metadata.tar and every part required by the chosen source/split. The complete root collection requires all four splits.

Dataset identity and splits

Collection: cofl-s-training-v1.
Revision: 20260907-v1.
Profile: ground_sector_v1.
Root collection.json SHA-256: cb3d7c23167ff32fa534e8aade35f3487ae5e084c5cb7b6c5780adc98ddb3b1b.

Source / split	Episodes	Observations	Annotations	Eligible annotations	Native shards
R2R / train	10,809	630,582	1,012,421	1,005,520	1,043
RxR / train	19,750	1,809,806	2,785,567	2,752,991	2,999
R2R / val_unseen	1,839	105,459	170,944	170,018	177
RxR / val_unseen	3,648	317,226	502,152	497,366	537

The complete training split has 3,758,511 eligible annotations. The released collection retains the original eligibility flags and all original annotation rows. The training recipe selects 8,192 validation samples from the complete val_unseen collection using its saved seed and subset identity; 8,192 is not the size of the validation dataset.

These are training-data splits. The separately defined R2R-1200 and RxR-1600 online benchmark subsets have their own manifests and preparation instructions.

Contents and storage

Native storage is Parquet plus Zarr v2. Observations and annotations remain separate, with stable references between them. The data includes RGB images, metric depth and validity masks, full and segment instructions, sector fields and supervision masks, action labels and geometric/trajectory metadata.

Keep the original relative layout when extracting archive parts:

datasets/cofl-s/
  collection.json
  r2r/
    train/collection.json
    train/<native-shard>/...
    val_unseen/collection.json
    val_unseen/<native-shard>/...
  rxr/
    train/collection.json
    train/<native-shard>/...
    val_unseen/collection.json
    val_unseen/<native-shard>/...

The source inventory is 352,221,117,480 bytes in 1,848,275 files. Four conversion_state.json migration work files (1,108,827 bytes total) are excluded; they are not needed by the reader. Planned archive payloads contain 352,220,008,653 bytes in 1,848,271 files, plus tar headers. Allow at least 360 GB for extracted data, and additional space for the archive parts you keep locally.

Archive packaging preserves original file bytes, manifests, sample IDs, annotation eligibility, split membership and ordering. Archive headers use relative paths and neutral ownership fields. The archive checksum list and part inventory will identify the bytes needed for each source/split.

Use with CoFL-S

Install the matching CoFL-S source release and its locked dependencies. Once the complete archives have been downloaded, checked and extracted, run from the source repository root:

uv run --locked cofl data validate datasets/cofl-s
uv run --locked cofl train --config configs/cofl_s_formal.yaml

The existing reader also opens the collection directly:

from cofl.data import open_dataset

dataset = open_dataset("datasets/cofl-s", split="train", include_extras=False)
sample = dataset[0]

The matching model is the separately distributed CoFL-S checkpoint. Its repository also contains the matching source-code snapshot. Dataset downloads do not include a model checkpoint or Habitat scene assets. Training from prepared data uses the stored observations and labels; online navigation evaluation additionally needs its benchmark annotations, licensed simulator scenes and a compatible Habitat environment.

Sources, license and access

Read LICENSE and THIRD_PARTY_NOTICES.md. The license statement covers CoFL's own contributions and identifies the independent terms of the source scenes and annotations. Dataset publication does not turn upstream material into unrestricted or commercially licensed data.

This public dataset repository requires agreement before automatic access. The publisher retains access-request records. Source-data terms continue to apply to downloaded and redistributed copies.

This collection preserves historical generated labels. A newer generation pipeline is not a claim of bit-for-bit regeneration of this exact collection.

Verification scope

The metadata audit covered 109,397 JSON/Zarr metadata files and all 14,268 Parquet files (7,370,203 rows), with no matches for the checked private-path and username patterns. It did not inspect image content or prove that no other form of personal information exists. A complete source manifest chain matches the checkpoint dataset identity.

A real native shard was packaged and extracted: all 389 file payloads were verified byte-for-byte, and five evenly spaced samples from its 1,020 samples were equal through the original reader. See the payload check and reader check. All 92 release parts have also been re-read and verified after packaging.

Downloads last month: 30

Total file size:

354 GB