RekaDaily-10k-raw · Dataset Card

RekaDaily-10k-raw: Dataset Card

Written by Reka AI, published under apache-2.0, revision 800e8ba00cf0, read 2026-09-18. Shown as written; SAVRN's own facts about this dataset are on its page.

Raw, unscripted, first-person daily-life video, collected through Claru, Reka's data collection marketplace — recorded by paid collectors in their own homes and workplaces on head-mounted and handheld phones, across multiple regions.

Videos are delivered as recorded — no cuts, no trimming, no editing, no filtering beyond basic integrity checks. A processed tier (short clips with machine captions) is released separately under the same RekaDaily-10k prefix.

This is the full RekaDaily-10k release: 10,865 hours / 412,050 videos / 11,127 shards / 80 TB. See the release announcement. Additional recordings may be appended over time; the metadata/ tables and this line are updated whenever that happens.

The Dataset Viewer above opens on the browse table: every video as a thumbnail next to its full metadata row, so the collection can be skimmed without downloading anything. The metadata config is the same fields without the images, for lighter programmatic reads.

Browse either release visually in the RekaDaily-10k Viewer Space: it plays sample videos and clips inline with their full metadata record beside them, and switches between the raw recordings and the processed clips from a dropdown at the top.

Structure

Videos are packed into WebDataset tar archives (~8 GB), organized by collection project:

data/<project>/shard-NNNNN.tar   # <video_id>.<mp4|mov> + <video_id>.json, plus <video_id>.imu.csv and
                                 # <video_id>.frames.csv for sessions recorded with motion sensors
metadata/browse.parquet          # one row per video: thumbnail + every metadata field
metadata/index.parquet           # the same metadata fields, without thumbnails

The metadata/ parquets are refreshed at the end of each upload batch, so they can trail the newest shards briefly while a batch is in flight.

Load the video shards with any WebDataset reader, e.g.:

import webdataset as wds
url = "https://huggingface.co/datasets/RekaAI/RekaDaily-10k-raw/resolve/main/data/egocentric_household_tasks/shard-00000.tar"
ds = wds.WebDataset(url)

The tar shards are not previewed in the Hub's Dataset Viewer (its WebDataset preview is currently broken platform-wide for video archives); they download and stream normally.

Projects: egocentric_household_tasks, egocentric_household_tasks_usa, residential_egocentric_latam_upload_via_claru, egocentric_commercial_environments, video_capture_activities, video_capture, video_capture_first_person_videos_phone, egocentric_household_tasks_imu, egocentric_household_tasks_imu_usa_canada, egocentric_household_tasks_imu_mexico. The three _imu projects ship motion-sensor sidecars (see below).

Metadata fields

Each .json sidecar carries the fields below; metadata/index.parquet and metadata/browse.parquet carry the same set, one row per video (browse adds the thumbnail image column):

field	description
`video_id`	unique id, matches the media file name
`project`	Claru collection project
`flow`, `activities`	activity taxonomy (activities-type projects): session scenario + performed actions
`category`, `subcategory`	category taxonomy (video-capture-type projects)
`lighting`	lighting condition, where captured
`duration_s`, `fps`, `width`, `height`, `num_frames`, `codec`	probe stats
`collector`	salted-hash collector id — distinct values ≈ distinct environments
`has_imu`	`true` when the video ships with `.imu.csv` and `.frames.csv`

Each video populates one taxonomy family (flow/activities or category/subcategory) depending on its project type.

IMU

Sessions from the newer collection projects were recorded with the phone's motion sensors alongside the camera. For those videos each shard carries two extra files with the same stem as the video:

file	contents
`<video_id>.imu.csv`	one row per sample: user acceleration with gravity removed (g), gravity vector (g), bias-corrected rotation rate (rad/s), attitude as quaternion and rotation matrix, plus raw accelerometer (g), gyroscope (rad/s) and magnetometer (µT) tracks, each on its own hardware timestamp column
`<video_id>.frames.csv`	presentation timestamp of every video frame on the same clock as the IMU, for frame-accurate joins

All timestamps are on one monotonic device clock per session. The .json sidecar for these videos additionally carries an imu block with the sample rates and frame anchor (sync), camera intrinsics and distortion coefficients with a provenance tag (intrinsics), the camera-to-IMU rotation (extrinsics.cam_to_imu; translation is null where it has not been physically measured), and the delivered frame geometry (resolution, including any rotation baked into the pixels). Attitude is gravity-aligned with arbitrary yaw; the magnetometer is provided raw and is not fused into attitude.

Videos without motion sensors have no .imu.csv/.frames.csv entries and their .json omits the imu block. Readers that pick only the mp4/json keys keep working unchanged.

Consent, privacy & takedown

This dataset was collected through Claru, Reka's data collection marketplace, as described in the release announcement. Collectors are paid contractors who opt in, and every session is recorded with the wearer's knowledge and agreement. Collectors are instructed to record only with the agreement of other adults present and to keep others out of frame where that is not possible.

Every video in this release has been processed to remove container metadata — GPS coordinates, device identifiers, and capture timestamps — and verified clean before upload, in addition to the automated PII screening described in the announcement. Screening is not perfect. If you find something in this release that should not be there, tell us and we will remove it: contact [email protected].

License

Apache 2.0 — use, redistribute, and build on this data, including commercially, with attribution per the license terms.