{
  "slug": "anymo-bench-2026",
  "title": "AnyMo Bench",
  "aliases": [
    "AnyMo-Bench",
    "AnyMo Bench Fine150",
    "AnyMo Bench Core50"
  ],
  "canonical_source_url": "https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench",
  "project_url": "https://baiyuchen.com/project/AnyMo",
  "introducing_paper_url": "https://arxiv.org/abs/2605.22715v2",
  "source_dataset": {
    "name": "Nymeria",
    "paper_venue": "ECCV 2024",
    "repository_url": "https://github.com/facebookresearch/nymeria_dataset"
  },
  "paper_status": "arXiv v2 preprint submitted 2026-05-21 and revised 2026-05-23; no peer-reviewed venue identified",
  "dataset_type": "real-world wearable inertial-measurement-unit activity-recognition benchmark derived from Nymeria",
  "domain": "in-the-wild human motion and wearable sensing",
  "release_year": 2026,
  "repository_snapshot": {
    "url": "https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench",
    "commit": "2316ef6d3d8544564a55d0cf682aee03d77c2f42",
    "commit_authored_at": "2026-05-29T18:58:33Z",
    "license_frontmatter": "CC BY-NC 4.0"
  },
  "population": {
    "subjects": 196,
    "train_subjects": 157,
    "test_subjects": 39,
    "subject_split": "subject-disjoint 8:2 split with seed 42"
  },
  "statistics": {
    "paper_and_card_headline_windows": 154695,
    "exported_rows_per_configuration": 154839,
    "train_rows_per_configuration": 123874,
    "test_rows_per_configuration": 30965,
    "reported_hours": 211.6,
    "configurations": 4,
    "parquet_shards_per_configuration": {
      "train": 25,
      "test": 7
    },
    "sampling_rate_hz": 60,
    "min_timesteps": 60,
    "max_timesteps": 300,
    "channels": 18
  },
  "configurations": [
    {
      "name": "AnyMo-Bench-150-US",
      "classes": 150,
      "evaluation": "unseen subject",
      "train_device": "first co-located IMU",
      "test_device": "first co-located IMU"
    },
    {
      "name": "AnyMo-Bench-150-USCD",
      "classes": 150,
      "evaluation": "unseen subject plus cross device",
      "train_device": "first co-located IMU",
      "test_device": "second co-located IMU"
    },
    {
      "name": "AnyMo-Bench-50-US",
      "classes": 50,
      "evaluation": "unseen subject",
      "train_device": "first co-located IMU",
      "test_device": "first co-located IMU"
    },
    {
      "name": "AnyMo-Bench-50-USCD",
      "classes": 50,
      "evaluation": "unseen subject plus cross device",
      "train_device": "first co-located IMU",
      "test_device": "second co-located IMU"
    }
  ],
  "body_positions": [
    "head",
    "left wrist",
    "right wrist"
  ],
  "channel_schema": [
    "three-axis accelerometer and three-axis gyroscope from the head",
    "three-axis accelerometer and three-axis gyroscope from the left wrist",
    "three-axis accelerometer and three-axis gyroscope from the right wrist"
  ],
  "tasks": [
    "fine-grained human activity recognition",
    "unseen-subject transfer",
    "same-position cross-device transfer",
    "long-tail classification"
  ],
  "baseline_protocol": {
    "models": [
      "DeepConvLSTM",
      "MantisV2",
      "COMODO with a MantisV2 IMU backbone and TimeSformer video backbone"
    ],
    "metrics": [
      "Acc@1",
      "Acc@5",
      "macro-F1"
    ],
    "fixed_length_handling": "DeepConvLSTM pads short windows and splits long windows; MantisV2 and COMODO interpolate to their model input length"
  },
  "row_count_tension": "The paper and dataset card state 154,695 windows, but the published split counts and official dataset_info.json contain 123,874 train plus 30,965 test rows, or 154,839 rows per configuration. The exported-row counts are treated as artifact-grounded while the headline value is preserved as reported.",
  "known_limitations": [
    "all four configurations reuse the same underlying windows under two label taxonomies and two device protocols rather than constituting four independent corpora",
    "the benchmark is derived from one source dataset and three body positions",
    "labels are produced by a multi-stage pipeline combining language-model proposals and adjudication, embedding-based consolidation, and human review",
    "Fine150 and Core50 remain long-tailed and some visually or semantically dependent labels were excluded",
    "cross-device tests change the co-located IMU unit while keeping body position fixed; they do not cover arbitrary new device families, placements, or sampling protocols",
    "the paper/card headline row count disagrees with the published split and metadata counts",
    "no actions, control inputs, rewards, interventions, or counterfactual protocol are included"
  ],
  "license_note": "Hugging Face frontmatter declares CC BY-NC 4.0, and the card says the benchmark follows Nymeria non-commercial research-use terms. Reuse should pin both the AnyMo Bench revision and applicable Nymeria terms.",
  "access_note": "Public Hugging Face Parquet release; this knowledge base stores metadata and documentation only and does not mirror dataset payloads.",
  "created": "2026-08-10",
  "updated": "2026-08-10"
}