# AnyMo Bench dataset source record

## Canonical sources

- Dataset repository: <https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench>
- Introducing paper: <https://arxiv.org/abs/2605.22715v2>
- Project page: <https://baiyuchen.com/project/AnyMo>
- Upstream Nymeria tools: <https://github.com/facebookresearch/nymeria_dataset>
- Pinned dataset revision: `2316ef6d3d8544564a55d0cf682aee03d77c2f42` (2026-05-29)

This directory stores metadata and documentation only. Dataset payloads are not mirrored.

## Preserved metadata artifacts

- `official_dataset_card.md`: exact README/data-card snapshot from the pinned Hugging Face Git revision.
- `official_dataset_info.json`: official build and split statistics from the pinned revision.
- `fine150_labels.json`: official Fine150 label vocabulary.
- `core50_labels.json`: official Core50 label vocabulary.
- `split_subjects.json`: official subject-disjoint split membership.
- `metadata.json`: normalized knowledge-base metadata and caveats.

## Dataset summary

AnyMo Bench is a fine-grained wearable IMU human-activity-recognition benchmark derived from synchronized real Nymeria streams. It targets two transfer axes:

1. **Unseen subjects:** train and test subjects are disjoint.
2. **Cross device at fixed placement:** training uses the first IMU at each body position, while testing uses the second co-located IMU at the same position.

The underlying source has 196 subjects and three observed body positions: head, left wrist, and right wrist. The subject-disjoint split uses seed 42 and contains 157 training subjects and 39 test subjects.

## Configurations

| Configuration | Label space | Transfer setting | Train IMU | Test IMU |
|---|---:|---|---|---|
| `AnyMo-Bench-150-US` | Fine150 | unseen subject | first unit | first unit |
| `AnyMo-Bench-150-USCD` | Fine150 | unseen subject + cross device | first unit | second co-located unit |
| `AnyMo-Bench-50-US` | Core50 | unseen subject | first unit | first unit |
| `AnyMo-Bench-50-USCD` | Core50 | unseen subject + cross device | first unit | second co-located unit |

The four configurations are different task/protocol views over the same window inventory, not four independent datasets.

## Observation schema

Each row contains:

- `sample_id`, `session_id`, and anonymized `subject_id`;
- the combined body-position group and selected device identifier;
- interval boundaries in seconds and Nymeria global nanoseconds;
- an activity label and zero-based label identifier;
- an `imu` array with shape `[T, 18]`, where the release metadata reports `60 <= T <= 300`.

The 18 numeric features concatenate three-axis accelerometer and three-axis gyroscope values from head, left wrist, and right wrist. Streams are synchronized to 60 Hz. Exported short windows retain variable length; long source windows are split into non-overlapping windows of at most 300 timesteps.

## Label construction

The paper derives AnyMo Bench from the larger AnyMo-180 vocabulary:

- language models propose and normalize activity labels from Nymeria atomic-action text;
- embedding-based clustering consolidates semantically close names;
- human experts review candidate labels;
- repeated enum-label assignments and adjudication remove or relabel uncertain rows;
- Fine150 removes unstable, sparse, or context-dependent categories;
- Core50 merges Fine150 categories using motion similarity, sample support, and interpretability.

This is a mixed automated/human curation pipeline. It gives broader fine-grained coverage than many wearable HAR benchmarks but remains dependent on the annotation models, consolidation policy, and IMU-observability judgments.

## Published split and row-count tension

The paper and dataset card both state that the benchmark contains **154,695** activity windows. The same sources report **123,874** training rows and **30,965** test rows per configuration. Those split counts sum to **154,839**, and the pinned `official_dataset_info.json` confirms 154,839 exported rows per configuration.

The knowledge base therefore preserves both values:

- `154,695`: paper/card headline claim;
- `154,839`: artifact-grounded exported-row count per configuration.

The four configurations expose 619,356 rows in aggregate because the same 154,839 windows are repeated under two taxonomies and two device protocols.

## Baseline protocol

The paper reports DeepConvLSTM, MantisV2, and COMODO baselines with Acc@1, Acc@5, and macro-F1. DeepConvLSTM and MantisV2 are trained for 100 epochs on the benchmark split; COMODO is adapted for 20 epochs with MantisV2 and TimeSformer backbones.

Headline difficulty is concentrated in same-position cross-device transfer. For example, on Core50 the best reported unseen-subject Acc@1 is 46.2, while the strongest cross-device Acc@1 is 32.6. These are paper-reported baselines rather than independently reproduced scores.

## License and access

Hugging Face frontmatter declares **CC BY-NC 4.0**. The data card additionally says that AnyMo Bench follows Nymeria non-commercial research-use terms. Reuse should pin this dataset revision and inspect the applicable Nymeria terms rather than assuming the paper's CC BY 4.0 license governs the data.

The Parquet payload is publicly hosted on Hugging Face. This knowledge base intentionally stores only metadata, vocabularies, split membership, and documentation.

## Limitations

- One upstream corpus and three body positions limit setup diversity.
- Cross-device evaluation changes a co-located unit, not arbitrary placement or device family.
- Fine150 and Core50 remain long-tailed.
- Automatic annotation and semantic consolidation can introduce label noise or ontology bias despite human review.
- The public row-count metadata is internally inconsistent by 144 windows.
- No actions, control inputs, interventions, rewards, or counterfactual evaluation are provided.