# Action100M Dataset Metadata Snapshot

Fetched: 2026-07-15

## Canonical Sources

- Official Hugging Face preview: https://huggingface.co/datasets/facebook/action100m-preview
- Official GitHub repository: https://github.com/facebookresearch/Action100M
- Paper: https://arxiv.org/abs/2601.10592
- Accepted workshop paper: https://openaccess.thecvf.com/content/CVPR2026W/EgoVis/html/Chen_Action100M_A_Large-scale_Video_Action_Dataset_CVPRW_2026_paper.html
- Official author announcement: https://x.com/Delong0_0/status/2012127335647002713

## What Is Public

The Hugging Face repository publishes a **10% preview by source-video rows** of Action100M:

- 120,000 rows in one `train` split;
- one row per source video;
- 120 Parquet shards;
- 39,173,609,923 compressed/download bytes according to the Dataset Server;
- 86,153,267,321 decoded dataset bytes according to the Dataset Server.

The preview schema contains `video_uid`, `metadata`, and `nodes`. It has no video-binary field. It is an annotation/metadata release keyed by YouTube video IDs, not a mirror of video payloads.

## Full-Dataset Scale Reported By The Paper

- 1,199,096 face-blurred HowTo100M videos;
- approximately 14.6 years of video duration;
- 147,092,653 hierarchical temporal segments;
- approximately 21.27 billion annotation words;
- approximately 205 GB to store annotations, metadata, and the complete Trees of Captions;
- ASR transcripts retrieved for 72% of source videos;
- 3.23% of segments receive an `N/A` action label;
- duration distribution: 64% at 0-3 seconds, 23.8% at 3-10 seconds, 10.2% at 10-60 seconds, and about 2% above one minute.

## Row And Node Schema

Each video row contains:

- `video_uid`: YouTube video ID;
- `metadata`: title, description, view count, like count, duration, upload date, and timestamped ASR transcript entries;
- `nodes`: the hierarchical Tree-of-Captions.

Each node contains hierarchy and temporal fields, plus nullable model outputs:

- `node_id`, `parent_id`, and `level` for hierarchy;
- `start` and `end` boundaries in seconds;
- `plm_action` and `plm_caption` from PerceptionLM-3B;
- `llama3_caption` from the middle frame of leaf nodes;
- optional `gpt.summary.brief` and `gpt.summary.detailed`;
- optional `gpt.action.brief`, `gpt.action.detailed`, and `gpt.action.actor`.

The Action100M paper says nodes shorter than four seconds are discarded from the GPT-OSS aggregation stage, but they remain in the released Tree-of-Captions rows. In the official `hySSAAw4t24.json` sample, all 639 of 917 nodes shorter than four seconds have `gpt: null`; 262 of 278 nodes at least four seconds long have non-empty GPT fields and 16 still have null GPT output. This is a one-video sample audit, not a full-corpus missingness estimate.

The GPT fields are generated by GPT-OSS-120B through three rounds of Self-Refine from local/global Tree-of-Captions context, video metadata, and ASR. They are automatic pseudo-labels, not human annotations.

## Relationship To VLWM

The Action100M paper explicitly states that its pipeline extends and improves the data-generation procedure introduced in the VLWM paper. Action100M is not the exact VLWM training dataset: it is HowTo100M-only and exposes segment summaries/actions/actors, while VLWM reports six video sources plus NaturalReasoning and trains on goals, goal interpretations, and explicit textual action-conditioned state changes.

## Access And License

The preview is public and ungated under the FAIR Noncommercial Research License. Commercial use is prohibited. The source videos remain external YouTube/HowTo100M assets; users must separately handle video availability, rights, and upstream terms.