khora-vie/khora-validation-data-public

Khora validation data (public)

The public part of the validation set for the Khora AI layer: the EgoDex and EgoVerse egocentric clips that every processor (pose, captions, retargeting) is tested on, plus the EgoVerse episodes used to train Khora's L2 sub-goal captioning model. The folder layout mirrors data/raw/ of the Khora repo, so one download puts every clip where ai/eval/datasets/manifest.yaml expects it. The in-house phone recordings are in a separate private dataset.

Contents

Folder	Source	Clips	Video	Size	Format
egodex_lerobot/test/<task>/	EgoDex (Apple), test split, LeRobot v3.0 copy by griffinlabs	52 tasks, one clip each	33 min (manifest clips)	6.7 GB	LeRobot v3.0 folders (AV1 video, parquet poses incl. ARKit hands and camera); each folder holds all test episodes of its task
egoverse/processed_v3/{aria,mecka,scale}/	EgoVerse (GaTech RL2), rldb bucket	52	103 min	9.8 GB	one zarr per episode (JPEG frames, hand keypoints, camera pose, language annotations)
egoverse_train/	EgoVerse Scale episodes of 2026-05-02	999 episodes, 1,472 windows	11.2 h	13 GB	L2 training windows, see below

EgoDex: 52 different tasks, one manifest clip each (no task repeated), to match EgoVerse's 52 clips: 10 picked on 2026-10-09 (5 longest-task episodes, 5 median ones), the other 42 the longest test episode of each task, across kitchen, cleaning, cloth/paper/string, assembly/tools, games and office tasks.

egoverse_train/: L2 captioning training data

The data the L2 model (qwen36-27b-l2-scale-vision12, checkpoint 350) was trained on: 1,000 EgoVerse Scale episodes recorded on 2026-05-02 (episodes_scale_2026-05-02_seed0.txt), cut into windows of 45-64 s along the human sub-goal spans, never through a span.

Path	What
videos/<episode>_<start s>.mp4	one H.264 window (source fps, short side <= 540 px)
manifests/l2_scale_train.jsonl	1,433 windows from 969 episodes (10.9 h)
manifests/l2_scale_val.jsonl	39 windows from 30 other episodes (0.3 h)
episodes_scale_2026-05-02_seed0.txt	the 1,000 episodes downloaded (zarr paths in the rldb bucket)

Each manifest row: video (relative to the dataset root), system (the L2 instruction), user (task and window length), target (the window's human sub-goals, one MM:SS.d - MM:SS.d: <sub-goal> line each, timed from the window start). None of these episodes is among the validation clips above (those are from 2026-04-30 and 2026-05-01).

Sources

• EgoDex: Apple, https://github.com/apple-aiml-research/ml-egodex (CC-BY-NC-ND); copy https://huggingface.co/datasets/griffinlabs/EgoDex-LeRobot-v3.0, verified against Apple's test.zip.
• EgoVerse: GaTech RL2, https://github.com/GaTech-RL2/EgoVerse.

Download

export HF_HOME=/scratch/$USER/.cache/huggingface     # keep the cache off your home quota
hf download khora-vie/khora-validation-data-public --repo-type dataset --local-dir /scratch/$USER/khora/data/raw

Part of it: --include "egodex_lerobot/*" (or "egoverse/*", "egoverse_train/*").

How it maps to the Khora repo

• ai/eval/datasets/manifest.yaml has one entry per clip (clip_id, source, original_id, storage_uri), and roots: maps each source to the folders above.
• ai/eval/datasets/adapters/<source>.py converts a clip into the Khora clip schema (schemas/clip.schema.json).
• Clip ids: egodex-test-<task>-ep<N>, egoverse-<aria|mecka|scale>-<episode id>.