This repository bundles five action-progress training sets whose sources carry different terms: DROID (CC-BY 4.0), Ego4D (Ego4D licence agreement required; no redistribution), HOIGen-1M (its terms), Penn Action and the TCC pouring data (original terms; not for redistribution). Request access only for research use, state your Ego4D licence if you use the Ego4D folder, and respect each source's terms. Access is granted manually.

Action progress: five video datasets (s31)

Training data of Whose Time Is It? Instance-Specific Action-Progress Control for Video Editing (LTX-2.5 22B actor-progress LoRA, Gerard Donahue's RIFE-16 method). Each folder is one training set with its own card (<folder>/README.md) describing the canvas, frames, progress source, masks, captions and the source's terms. No VAE latents or RIFE videos are included; they are rebuilt from the source videos with the repository's preprocessing.

Access. This repository bundles five action-progress training sets whose sources carry different terms: DROID (CC-BY 4.0), Ego4D (Ego4D licence agreement required; no redistribution), HOIGen-1M (its terms), Penn Action and the TCC pouring data (original terms; not for redistribution). Request access only for research use, state your Ego4D licence if you use the Ego4D folder, and respect each source's terms. Access is granted manually.

folder	source	progress	canvas	training clips	hold-outs	size
`droid/`	DROID (robot)	ground truth (trajectory phase)	960x544	3,193	4	16.05 GB
`ego4d/`	Ego4D (egocentric)	VLM-measured	896x672	16,750	4	102.86 GB
`hoi-multi/`	HOIGen-1M (2 to 4 actors)	VLM-measured per actor	960x544	1,479	4	2.64 GB
`penn-action/`	Penn Action (sports)	ground truth (phase labels)	928x512	308	4	0.82 GB
`pouring/`	TCC pouring	ground truth (subaction labels)	512x928	181	4	0.14 GB

Every folder has the same layout: annotations.jsonl (one record per training clip with the window, natural progress curve, the four cached variants, the LTX-style caption in caption_ltx, the original task and the SAM 3.1 phrases), holdouts.jsonl (validation clips), masks/part-*.zip (packed actor and object masks), videos/part-*.zip (native source videos, byte for byte), configs/, and for Penn Action and pouring selection_report.json.

Evaluation sets

Each of the four folders below has an eval/ set of at least 200 clips that none of the released runs trained on (pouring has none: its validation split has 10 videos). Every clip carries the LTX caption used as prompt, the first frame and actor occupancy on the training canvas, the real clip on that canvas, the SAM 3.1 masks, the native source video byte for byte, and the progress curves a model is asked to follow (native, early, late; Ego4D also native_gt, early_gt, late_gt from the FHO critical frames). <folder>/eval/README.md describes how the clips were chosen; load the records with the <folder>-eval config (split test).

folder	clips from	progress	clips	size
`droid/eval/`	DROID episodes outside the training subset and hold-outs (57+ frames)	ground truth	200	1.54 GB
`ego4d/eval/`	official Ego4D val split (no run trained on it)	VLM curve and FHO ground truth	200	2.63 GB
`hoi-multi/eval/`	HOIGen-1M clips with 2 to 4 actors outside every multi-actor subset	VLM-measured per actor	200	1.21 GB
`penn-action/eval/`	official Penn Action test split, several windows per clip	ground truth	200	1.21 GB

hf download sebastian-stapf/action-progress-video --repo-type dataset --include "ego4d/eval/*" --local-dir action-progress

Download a single folder:

hf download sebastian-stapf/action-progress-video --repo-type dataset --include "droid/*" --local-dir action-progress

Downloads last month: 52

Total file size:

129 GB