You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

AffectSim

This publicly visible, manually gated preview contains metadata and privacy-clean Episode YAMLs for all 27,647 embodied emotion-perception episodes. NOKOV motion PKLs are included; Emilya and KDAE motion assets remain external and are not redistributed.

Downloads require the owner's manual approval and an authenticated Hugging Face account. Public visibility does not make downloads unrestricted or grant new rights to upstream assets. The original-name archive below retains actor codes and test-set emotion information; this repository is not a blind-test package.

Download original-name Episode YAMLs

The original-name V6 YAML ZIP contains 27,647 configurations with their original YAML filenames, PKL filenames, directory hierarchy, and relative references intact. Only local absolute-path fields (asset hints and generator provenance) were removed. No private mapping files are included. See the original-name download guide.

This separate archive is not anonymous or test-label-hidden: original actor codes and emotion-bearing names/metadata are preserved. Its download is manually gated; it contains no motion or scene binaries and preserves the historical configuration schemas, including the original NOKOV placement-record format. It is not a drop-in replacement for the portable bundle below.

Download portable anonymous Episode YAMLs

The V6 configuration ZIP contains all 27,647 portable YAMLs: 17,901 Emilya, 8,306 KDAE, and 1,440 NOKOV. It contains only YAML files, without private ID mappings, source-path records, motion PKLs, BVHs, scenes, or videos. Test target labels remain hidden. See download, checksum, extraction, and loading instructions. Access requires manual approval for this gated dataset.

Current contents

Split Episodes Labels in standard metadata Released passive media
train 19,662 five-class labels included P-Init + P-Ref + P-Mon: 58,986 MP4s
validation 4,044 five-class labels included P-Init + P-Ref + P-Mon: 12,132 MP4s
test 3,941 hidden; all values are null P-Init + P-Ref + P-Mon: 11,823 MP4s
Source Episodes
Emilya 17,901
KDAE 8,306
NOKOV dyadic 1,440
Released asset Availability
Emilya Episode YAML 17,901 files; train 11,999, validation 3,003, test 2,899
KDAE Episode YAML 8,306 files; train 6,643, validation 831, test 832
NOKOV Episode YAML 1,440 files; train 1,020, validation 210, test 210
NOKOV motion PKL 96 files; 48 synchronized pairs with two actors each
Emilya/KDAE motion PKL not included; local, matching motions are required
Emilya grounding support 94 anonymous JSON profiles used by 1,791 episodes
V6 train P-Init/P-Mon videos 39,324 MP4s; 23,414,739,139 bytes
V6 validation P-Init/P-Mon videos 8,088 MP4s; 4,258,403,299 bytes
V6 test P-Init/P-Ref/P-Mon videos 11,823 MP4s; 28,459,262,640 bytes
V6 validation P-Ref videos 4,044 MP4s; 13,974,649,559 bytes
V6 train P-Ref videos 19,662 MP4s; 78,168,617,635 bytes
A-Obs videos not uploaded
HSSD scene binaries obtain from the official HSSD-Hab release

All three passive views are included for each of the 27,647 episodes: 82,941 MP4s, totaling 148,275,672,272 bytes. The videos retain their original encoding; consolidation does not re-encode them.

The five labels are angry, fearful, happy, neutral, and sad. In the standard portable files, episode and motion IDs are opaque HMAC-derived identifiers. The secret salt and the source-ID mapping are retained offline and are not part of this repository.

Train and validation media is classified by split, view, source, and a two-hex shard: media/<split>/<p_init|p_mon|p_ref>/<source>/<shard>/<opaque-id>.mp4. P-Ref is included at media/train/p_ref/ and media/validation/p_ref/. Test media retains its existing unsharded layout.

Load

from datasets import load_dataset

dataset = load_dataset("wangzhilong123/AffectSim", token=True)
print(dataset)

Assets and licenses

The 57 HSSD scene IDs and episode counts are listed in hssd_57_scenes.csv. Use the clickable HSSD_57_SCENES.md table when browsing. Scene binaries are not mirrored here.

The released NOKOV configurations are indexed by nokov_episode_assets.csv.gz. NOKOV PKLs contain only the numeric Habitat pose payload required at runtime; participant identifiers and local source paths were removed without changing the motion arrays.

See DATA_ACCESS_AND_LICENSE.md and THIRD_PARTY_NOTICES.md before using this manually gated preview.

For Emilya/KDAE configuration downloads and local motion binding, see SINGLE_PERSON_EPISODES.md. The configurations preserve the frozen scene, character and robot placements, and playback parameters. Anonymous motion references prevent test labels from being exposed through original filenames. Downloading a YAML alone does not provide its external scene or motion assets.

Privacy and release boundary

The following guarantees apply to the standard portable files and the anonymous YAML bundle, not to the separately requested original-name ZIP. That ZIP preserves source actor codes and emotion information, including test-set information; access to this repository must therefore not be treated as access to a label-hidden test-only release. Private ID mapping files remain excluded from both variants.

  • no Windows, Linux-server, or user-home absolute paths in public manifests;
  • no participant identifiers or label-bearing source episode IDs;
  • only privacy-clean NOKOV motion PKLs are included; Emilya/KDAE PKLs are absent;
  • no HSSD mesh, NavMesh, avatar, or robot binary;
  • all released media uses opaque episode IDs;
  • no test emotion labels;
  • no A-Obs media;
  • P-Init, P-Ref, and P-Mon are included for every split.

The current consolidation audit is in training_reference_consolidation.json. Earlier audit files describe their respective historical release stages and have not been rewritten. The original audit is in audits/release_audit.json, the dedicated train/validation media audit is in audits/train_validation_nonref_media_release.json; the subsequent validation P-Ref addition is audited in audits/validation_reference_media_release.json. File hashes are in checksums/SHA256SUMS.txt. The single-person YAML lineage and runtime-field audit is in single_person_episode_asset_release.json.

Downloads last month
35