{
  "slug": "action100m-2026",
  "title": "Action100M: A Large-scale Video Action Dataset",
  "fetched_at": "2026-07-15",
  "official_dataset_url": "https://huggingface.co/datasets/facebook/action100m-preview",
  "official_dataset_revision": "e94f42ab16ae84c659417d60661a9e6dcc42336d",
  "official_dataset_last_modified": "2026-01-15T16:30:30Z",
  "official_code_url": "https://github.com/facebookresearch/Action100M",
  "paper_url": "https://arxiv.org/abs/2601.10592",
  "venue_url": "https://openaccess.thecvf.com/content/CVPR2026W/EgoVis/html/Chen_Action100M_A_Large-scale_Video_Action_Dataset_CVPRW_2026_paper.html",
  "venue": "CVPR 2026 Workshops, EgoVis",
  "license": "FAIR Noncommercial Research License v1 (2024-10-16)",
  "release_scope": {
    "public_fraction": "10% preview by source-video rows",
    "public_fraction_caveat": "not established as exactly 10% of hierarchical segments or decoded bytes",
    "rows": 120000,
    "split": "train",
    "parquet_shards": 120,
    "download_size_bytes": 39173609923,
    "decoded_dataset_size_bytes": 86153267321,
    "video_payload_in_schema": false
  },
  "full_dataset_reported": {
    "source": "HowTo100M instructional videos",
    "videos": 1199096,
    "duration_years": 14.6,
    "annotated_segments": 147092653,
    "annotation_words_billions": 21.27,
    "annotation_storage_gb": 205,
    "asr_transcript_coverage_percent": 72,
    "na_action_label_percent": 3.23,
    "segment_duration_distribution_percent": {
      "0_to_3_seconds": 64.0,
      "3_to_10_seconds": 23.8,
      "10_to_60_seconds": 10.2,
      "over_60_seconds": 2.0
    }
  },
  "row_schema": {
    "video_uid": "YouTube video id",
    "metadata": {
      "title": "string",
      "description": "string",
      "view_count": "int64",
      "like_count": "int64",
      "duration": "int64",
      "upload_date": "string",
      "transcript": "list of {time, text}"
    },
    "nodes": {
      "hierarchy": ["node_id", "parent_id", "level"],
      "temporal_bounds": ["start", "end"],
      "perception_fields": ["plm_action", "plm_caption", "llama3_caption"],
      "gpt_summary_fields": ["brief", "detailed"],
      "gpt_action_fields": ["brief", "detailed", "actor"],
      "gpt_nullable": true
    }
  },
  "official_sample_audit": {
    "sample_file": "https://raw.githubusercontent.com/facebookresearch/Action100M/main/data/hySSAAw4t24.json",
    "nodes": 917,
    "nodes_under_4_seconds": 639,
    "nodes_under_4_seconds_with_nonempty_gpt": 0,
    "nodes_at_least_4_seconds": 278,
    "nodes_at_least_4_seconds_with_nonempty_gpt": 262,
    "nodes_at_least_4_seconds_with_null_gpt": 16,
    "scope_note": "one official sample video; not a corpus-wide missingness estimate"
  },
  "annotation_pipeline": [
    "hierarchical temporal segmentation with V-JEPA 2 embeddings and agglomerative clustering",
    "segment captions with PerceptionLM-3B and leaf middle-frame captions with Llama-3.2-Vision-11B",
    "five structured fields from GPT-OSS-120B with three rounds of Self-Refine"
  ],
  "annotation_type": "fully automated open-vocabulary pseudo-labels",
  "relationship_to_vlwm": "The Action100M paper says its pipeline extends and improves the VLWM data-generation procedure, but the released schema and source mixture differ from the unreleased VLWM training corpus."
}