ELMA: Episodic Long-term Motion Dataset with Multi-segment Annotations

ELMA is a benchmark dataset for long-term text-to-motion generation with cross-segment anaphoric dependencies. Each episode is a 30-second human motion sequence divided into six consecutive 5-second segments, paired with segment-level captions and dependency annotations that mark when a later segment refers back to an earlier one (e.g. "continues jogging in place").

• Paper: ELMA: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation (NeurIPS 2026 Evaluations & Datasets Track)
• Code / baselines: https://github.com/shun31y/elma-benchmark
• Version: v2.0 — a corrected update of v1.0 (see What's new in v2.0).
• License: CC BY-NC 4.0

Reproducing the paper: the paper's experiments use v1.0, which stays available unchanged under the v1.0 tag. Pass --revision v1.0 to the download commands below.

Versions

Version	Tag	Notes
v2.0	v2.0 (current main)	Cleaner test set and root-height correction
v1.0	v1.0	Byte-identical to the data used for the experiments in the paper

What's new in v2.0

v2.0 keeps the v1.0 directory layout, file names, IDs and file formats, so code written for v1.0 runs on v2.0 without modification. The main changes are reduced obvious motion and caption noise in the test set and a global root height correction.

1. The v1.0 test set contained noisy motions and captions

ELMA motions are estimated from in-the-wild videos, so the v1.0 test set also contained clips whose estimated motion does not faithfully reproduce the source video (e.g., failed or unstable tracking), as well as captions that do not describe the motion. Generated motions were therefore being compared against noisy references, which makes evaluation numbers less reliable.

For v2.0, human reviewers compared each test clip's source video with its estimated SMPL-X motion side by side, and the test set was rebuilt from approved clips only, reducing obvious motion and caption noise.

2. Root height was not grounded

In v1.0, smpl_params_global.transl was taken directly from the motion estimator, and the body frequently floated above or sank below the ground plane. In v2.0, the vertical component of the global translation is corrected with a contact-aware, temporally smoothed floor offset estimated from foot contacts.

Summary of changes

	v1.0	v2.0
Episodes (raw/)	13,200	13,200 (same IDs)
train/	10,200	10,200 (identical IDs to v1.0)
test/	3,000	1,000 (human-approved subset of v1.0 test)
Excluded from both splits	–	2,000
transl / joints (global)	as estimated	floor-height corrected
smpl_params_global.transl_velocity	–	added
HML3D 263-D features / Mean / Std	v1.0	recomputed from corrected motion
HML3D anaphora test set	1,000	1,000
HML3D independent test set	1,000	not included
MotionStreamer272/	–	added (272-D MotionStreamer format)

• train is exactly the v1.0 train split. A model trained on v1.0 train has never seen any v2.0 test episode, because v2.0 test is a subset of v1.0 test.
• test contains 1,000 human-approved episodes only.
• The remaining 2,000 v1.0 test episodes are in neither split. They stay in raw/, texts/, HML3D/ and MotionStreamer272/ so that IDs and all.txt remain the same as v1.0, but they should not be used for training or evaluation.

Dataset summary

Episodes	13,200 (train 10,200 / test 1,000 / excluded 2,000)
Episode length	30 s = 6 segments × 5 s
Segment captions	79,200
Anaphora test subset	1,000 episodes (test_anaphora)
Motion representation	SMPL-X parameters (30 fps), HumanML3D features (20 fps), MotionStreamer 272-D features (30 fps)
Total size	~153 GB

Data source and collection

ELMA is constructed from publicly available online videos through a multi-stage pipeline: candidate videos were collected with motion-related search queries, segmented into scenes, filtered by duration and motion quality, and processed with a monocular human mesh recovery model to extract SMPL-X motion sequences. Captions and dependency annotations were generated per segment.

The original videos are not redistributed. This release contains only the processed motion data, captions and annotations. File names (C00000001, ...) are anonymized episode IDs.

Repository layout

The Hugging Face Hub allows at most 10,000 files per directory, so the per-episode files are distributed as uncompressed tar archives. Every archive stores dataset-relative paths, and extracting all of them into one directory reproduces the directory structure below exactly.

File	Extracts to	Size
HML3D.tar	HML3D/ (HumanML3D-format export used by the baselines)	~11 GB
MotionStreamer272.tar	MotionStreamer272/ (MotionStreamer 272-D export)	~13 GB
texts.tar	texts/	~27 MB
train-{00..10}-of-11.tar	train/ (1,000 episodes per shard)	~54 GB
test-00-of-01.tar	test/	~5 GB
raw-{00..13}-of-14.tar	raw/ (identical to train/ ∪ test/ ∪ excluded; optional)	~70 GB
meta/	stored as plain files	~0.1 GB
SHA256SUMS	checksums of every extracted file

Directory structure (after extraction)

ELMA/
├── raw/            13,200 × {id}.pt   SMPL-X motion for all episodes
├── train/          10,200 × {id}.pt   training episodes
├── test/            1,000 × {id}.pt   test episodes
├── texts/          13,200 × {id}.json segment captions with timestamps
├── meta/
│   ├── train.txt, test.txt            split episode IDs
│   ├── event_anaphora.jsonl           per-segment dependency annotations (79,200 rows)
│   └── test_anaphora.jsonl            1,000 anaphora evaluation targets
├── HML3D/          HumanML3D-format export used by the baselines
│   ├── new_joint_vecs/{id}.npy        (T, 263) features, 20 fps (+ 1,000 evaluation samples, see below)
│   ├── new_joints/{id}.npy            (T, 22, 3) joint positions
│   ├── texts/{id}.txt                 caption#POS-tags#start#end per segment (+ 1,000 evaluation samples)
│   ├── Mean.npy, Std.npy              normalization statistics (263,), computed on train
│   ├── train.txt, val.txt, test.txt, all.txt
│   ├── test_anaphora.txt
│   └── humanml_test_set_anaphora.json, humanml_test_set_anaphora_fullgen.json
└── MotionStreamer272/   MotionStreamer 272-D export
    ├── motion_data/{id}.npy           (900, 272) features, 30 fps
    ├── texts/{id}.txt                 caption#POS-tags#start#end per segment
    ├── mean_std/Mean.npy, Std.npy     normalization statistics (272,), computed on train
    └── split/train.txt, val.txt, test.txt, all.txt

val.txt is identical to test.txt, as in v1.0; there is no separate validation split. If you need one, hold out part of train.txt.

Besides the 13,200 episode files, HML3D/new_joint_vecs/ and HML3D/texts/ contain 1,000 evaluation samples named {id}__ana_segNN_XXXX (listed in test_anaphora.txt), where segNN is the target segment. They reuse the full motion of the source episode, so new_joints/ has no entries for them. Their generation settings are given in humanml_test_set_anaphora*.json.

Motion files (raw/, train/, test/)

Each .pt file is a dict loadable with torch.load(path, weights_only=False), containing 900 frames (30 s at 30 fps):

Key	Content
smpl_params_global	body_pose (900, 63), betas (900, 10), global_orient (900, 3), transl (900, 3), joints (900, 127, 3), transl_velocity (new in v2.0) in world coordinates
smpl_params_incam	the same SMPL-X parameters in camera coordinates (without joints)
K_fullimg	(900, 3, 3) camera intrinsics
net_outputs	raw outputs of the motion-recovery network

Captions (texts/{id}.json)

{
  "seq_name": "SEQ_C00004713",
  "frame_ann": {
    "labels": [
      {"proc_label": "The person sits on a mat twisting their upper body from side to side.",
       "start_t": 0.0, "end_t": 5.0, "act_cat": []},
      ...
    ]
  }
}

Dependency annotations (meta/*.jsonl)

One row per (episode, segment). Main fields:

Field	Meaning
episode_id, target_segment_id	segment being annotated (segments are 1-indexed)
target_caption	caption of the segment (matches texts/ in v2.0)
event_anaphora	whether the segment depends on an earlier segment
depends_on_segment_ids	antecedent segment IDs
dependency_type	event, state, object or none
anaphoric_expression_in_target, antecedent_expression	the linked expressions
babel_category, ...	coarse action category

Usage

Download everything (~153 GB) and extract:

hf download shun31y/ELMA --repo-type dataset --revision v2.0 --local-dir ELMA
cd ELMA
for f in *.tar; do tar -xf "$f" && rm "$f"; done
sha256sum -c --quiet SHA256SUMS   # optional: verify every extracted file

Download only the HumanML3D-format export used by the baselines (~11 GB):

hf download shun31y/ELMA HML3D.tar --repo-type dataset --revision v2.0 --local-dir ELMA
tar -xf ELMA/HML3D.tar -C ELMA

raw-*.tar duplicates train/ and test/ and also holds the excluded episodes; skip it with --exclude "raw-*" if you do not need the raw/ directory. Likewise, skip MotionStreamer272.tar if you do not use the 272-D format.

To use the export with the benchmark code, link it in and follow the code's README:

ln -s /path/to/ELMA/HML3D ./FlowMDM/dataset/HumanML3D

To reproduce the paper's results, download --revision v1.0 instead.

Notes for v1.0 users

• Re-run anything that depends on the test split or on normalization statistics. HML3D/Mean.npy/Std.npy and MotionStreamer272/mean_std/* are computed on the v2.0 train split from the corrected motion, so they differ from v1.0.
• The independent test set from v1.0 (HML3D/test_independent.txt, __ind_ files, meta/test_nonanaphora.jsonl) is not included in v2.0. Anaphora labels for every segment remain available in meta/event_anaphora.jsonl. Evaluation in v2.0 uses the anaphora test set (test_anaphora.txt, humanml_test_set_anaphora*.json).
• The target_caption field in meta/*.jsonl now matches the segment caption in texts/. (In v1.0 it could differ for some segments, because it held the caption at the time the anaphora label was assigned.)

Known issues

• ELMA was collected with an automatic pipeline, so annotations may contain noise.

Removal requests

If you appear in the source videos or own their rights and want an episode removed, please open a discussion on this dataset's Community tab or an issue on the GitHub repository. Removed episodes will be excluded from future versions.

License

The dataset is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). It is intended for non-commercial research use only.

Citation

@inproceedings{kato2026elma,
  title     = {{ELMA}: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation},
  author    = {Kato, Shun and Yang, Dixin and Amaya, Kotaro and Wada, Yuiga and Fukuda, Kazumi and
               Shibuya, Takashi and Mitsufuji, Yuki and Isogawa, Mariko},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  eprint    = {2610.13549},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}