SprongMotion-100K

101,331 human motion clips (132.2 hours) with English captions, generated with NVIDIA Kimodo-SOMA-RP-v1.1 from 82,677 prompts. Every clip is a skeletal animation of the SOMA-77 joints body at 30 fps, as joint rotations (Parquet) and as a BVH file. Each clip has one main caption, and many have hand-written alternative wordings and checked paraphrases (213,082 extra texts in all). Many of the written captions read like prose and name things the skeleton cannot show: see Caption style.

This is exactly the generated part of the training set of SprongMotion-14M 1.0, a small open text-to-motion model, with the same captions, caption fixes and drops. The model also trained on motion capture and hand-keyed game animation from other sources. Those are not repackaged here: the model card lists them.

By Seger Endendijk (segerend.nl). Sprong is Dutch for leap :)

License

The dataset is distributed under:

CC BY 4.0. Licensor, and the credit CC BY asks for: "Seger Endendijk (segerend.nl)".

Please do not use SprongMotion for military, warfare, nuclear or espionage purposes.

Intended uses

• Training and evaluating text-to-motion models.
• Motion retrieval and captioning.

Loading

from datasets import load_dataset
import numpy as np

dataset = load_dataset("SegerEnd/SprongMotion-100K", split="train")
clip = dataset[0]
print(clip["caption"])                      # e.g. "a man takes eight steps forward with a limp"
quats = np.array(clip["local_quats"])       # [T, 77, 4] local joint rotations, wxyz
root = np.array(clip["root_pos"])           # [T, 3] root (Hips) position in metres

The BVH files come as one zip per category. Unzipped in the dataset folder, each lands at its row's bvh path. One clip's BVH, straight from its zip:

import zipfile
from huggingface_hub import hf_hub_download

category = clip["bvh"].split("/")[1]
zip_path = hf_hub_download("SegerEnd/SprongMotion-100K", f"bvh/{category}.zip", repo_type="dataset")
bvh_text = zipfile.ZipFile(zip_path).read(clip["bvh"]).decode()

Structure

data/train-0000k-of-0000n.parquet       91,048 clips
data/validation-0000k-of-0000n.parquet  5,272 clips
data/test-0000k-of-0000n.parquet        5,011 clips
bvh/<category>.zip                      the category's BVH files, one per clip, as bvh/<category>/.../<id>.bvh
skeleton/soma77.json                    joint names, parents, rest pose

About 11 GB in all: 7.0 GB of Parquet and 4.1 GB of zipped BVH (101,331 files in 41 zips).

The validation and test clips are the ones SprongMotion-14M held out from training, chosen by a hash of the text, so nearly every prompt is in one split only (3 of 82,677 prompts and 41 of 82,352 captions have clips in two).

Benchmark overlap. Clips whose caption is a test prompt of NVIDIA's Kimodo Motion Generation Benchmark, and seed clips made from an action type in its test_content split, are not in this set (4,367 clips were left out), so models trained on it can be scored on that benchmark without having seen its test texts.

Columns

column	type	description
id	string	clip id, k<prompt_id>_<sample>
caption	string	what the clip shows: usually the prompt, fixed where the clip measurably does something else (see Checks)
prompt	string	the text the clip was generated from (for blocks clips: the block texts with their lengths)
user_prompts	list of string	hand-written other wordings of the same motion: short game-clip names, -ing forms, other names ("zigzag walk", "mummy walk")
paraphrases	list of string	rewordings of the caption, written by an LLM and checked so they keep the side, direction, count and gait
source	string	template, seed or written (see Prompts)
category	string	kind of motion (see Categories)
method	string	text, text+path (a root path constraint), text+pose keys (pose keyframes), blocks (several prompts in a row)
segments	list of {text, start, end}	for blocks clips: each block's text and frame range (end exclusive), empty otherwise
action	string	seed: SEED action type (e.g. jog_ff_start_180_R), others: null
scene	bool	the text needs a prop that carries the body (sitting, a box, stairs): the clip sits or steps onto an invisible prop, as in game animation
prompt_id	int32	prompt number, clips of the same prompt share it
duration_s	float32	length in seconds
frames	int32	number of frames (T)
fps	int16	30
flags	list of string	per-clip checks that fired at generation time (side, static, gait)
bvh	string	path of the clip's BVH file inside bvh/<category>.zip (and after unzipping in the dataset folder)
root_pos	float32 [T, 3]	root (Hips) position in metres, Y up
local_quats	float32 [T, 77, 4]	rotation of each joint relative to its parent, quaternion w, x, y, z

One row (root_pos and local_quats shortened):

{'id': 'k000009_00', 'caption': 'the person jogs in a circle, then stops and stands still',
 'prompt': 'the person jogs in a circle, then stops and stands still', 'source': 'template',
 'category': 'locomotion', 'action': None, 'scene': False, 'prompt_id': 9, 'duration_s': 4.6,
 'frames': 138, 'fps': 30, 'flags': [], 'user_prompts': [], 'paraphrases': [], 'method': 'text',
 'segments': [], 'bvh': 'bvh/locomotion/k000009_00.bvh',
 'root_pos': [[...], ...],     # 138 x 3
 'local_quats': [[...], ...]}  # 138 x 77 x 4

Skeleton and forward kinematics

skeleton/soma77.json holds the joint names, parents (-1 for the root) and rest_joints (rest pose joint positions, metres). Joint positions follow from:

from scipy.spatial.transform import Rotation as R

def fk(quats, root_pos, rest, parents):
    T, J = quats.shape[:2]
    local = R.from_quat(quats.reshape(-1, 4)[:, [1, 2, 3, 0]]).as_matrix().reshape(T, J, 3, 3)
    glob, pos = np.zeros_like(local), np.zeros((T, J, 3))
    for j, p in enumerate(parents):  # parents come before children
        if p < 0:
            glob[:, j], pos[:, j] = local[:, j], root_pos
        else:
            glob[:, j] = glob[:, p] @ local[:, j]
            pos[:, j] = pos[:, p] + glob[:, p] @ (rest[j] - rest[p])
    return pos  # [T, 77, 3] metres

This reproduces Kimodo's own joint positions to within 0.001 mm on every clip. The BVH files hold the same motion in centimetres (ZYX Euler channels). Read back, they match to within 0.062 mm.

SprongMotion-14M itself was trained on a 22-joint skeleton with 263-d motion features computed from these joints. That form is not included: it follows from the joint positions.

Prompts

source	clips	hours	what
seed	24,962	30.4	SEED Timeline Annotation texts (whole motions and timeline segments)
template	27,392	37.8	action lists combined with direction, speed, side, count, style and emotion
written	48,977	64.0	prompts written for breadth and detail: game moves, everyday actions, sports, dances, compositions ("walks backward while waving"), sequences ("A, then B"), characters ("walks like a mummy"), exact counts and turns, paths, long actions

Who wrote the prompts:

• About 34,817 of the written clips: prompts drafted by IBM Granite-3.3-2B-Instruct (Apache-2.0) from themes and concept lists the Licensor wrote, then filtered.
• The persona category (1,258 clips): written by Qwen3-4B-Instruct-2507 (Apache-2.0) playing different users of a text-to-motion tool (a typed prompt, kept as a user prompt, plus a plain meaning that the clip was generated from), then checked.
• seed: NVIDIA's SEED Timeline Annotations (see Provenance). Only the texts are used, the motion is generated for exactly that text.
• Other prompts and captions: written by the Licensor, or partly with the help of an AI assistant for variation.

Method

Kimodo-SOMA-RP-v1.1 ran with 50 DDIM steps, classifier-free guidance 2.0 on the text and its default post-processing.

method	clips	what
text	99,913	from the text alone
text+path	1,316	text plus a root path the body follows (paths, circles, long locomotion)
text+pose keys	82	text plus a few pose keyframes (gestures, transitions to the floor)
blocks	20	several prompts in a row (see segments)

Categories

category	clips	hours
seed	24,962	30.4
locomotion	8,578	14.3
sport	6,589	7.9
daily	6,334	8.1
combat	6,263	5.8
emote	4,482	5.0
dance	4,077	7.2
direction	3,974	5.7
gestures	3,722	3.3
move	3,640	4.7
side	3,376	5.2
work	1,912	2.7
social	1,751	2.1
interact	1,734	2.1
game	1,637	1.6
idle	1,531	2.6
sequence	1,454	3.2
fitness	1,435	2.0
count	1,388	1.9
star	1,287	1.8
persona	1,258	1.5
path	1,250	2.9
art	1,062	1.5
hit	955	0.7
hero	852	1.0
pose	726	1.0
compose	642	1.0
death	569	0.6
moves	553	0.6
ranged	537	0.6
evade	533	0.4
magic	515	0.6
precise	494	0.6
weak	372	0.6
character	290	0.4
clip	194	0.2
gap	179	0.2
long	104	0.3
transition	66	0.1
stylized	34	0.0
blocks	20	0.0

Durations

Clip lengths are 1.5-10.0 s (median 4.6 s).

length	clips
0-2 s	4,916
2-4 s	32,092
4-6 s	40,493
6-8 s	18,995
8-10 s	4,835

Checks

Applied to every clip before it went into SprongMotion-14M's training set. All counts below are clips left out or changed.

• Caption match: text-prompted clips that NVIDIA's TMR-SOMA-RP-v1 text-motion retrieval model (NVIDIA Open Model License, used only for this check) ranked far from their own prompt, and prompts whose samples mostly failed the measured checks, were left out.
• Review by eye: samples of every kind of prompt were watched as stick figures. Actions Kimodo does not really perform (push-ups that hold the pose, tying a shoelace, some named kicks and dances, ...) were left out, and the newer, smaller batches were judged clip by clip, with fast actions judged by per-frame measurement.
• Measured captions: the caption was fixed where the clip measurably does something else, or the clip was left out:
  • turns: the facing at the end against the start, rounded to 15 degrees. Clips that turn the named way get the angle in the caption ("then turns 90 degrees to the left"). Clips that barely turn or turn the other way are out. Clips captioned with a start facing that Kimodo cannot honour (it always starts facing forward) are out unless rewritten to the measured turn.
  • circles: path clips that do not close are captioned "curves all the way around", "circle" is kept for closed circles. Circle clips with the side swapped were recaptioned or left out.
  • counts: jumps, hops, claps and steps were counted and the caption fixed to the count.
  • stops: continuous actions that stand still for at least 1 s and a quarter of the clip end with ", then stops".
  • jog and run by measured speed: "jogs" only up to 2.0 m/s (median root speed), faster clips say "runs" (434 clips).
  • "lies down" only for clips that go down (head above 1.0 m at the start, below 0.5 m at the end). Clips that already lie at the start say "lies on the floor" (179 clips).
  • in place: clips that stay in place are not captioned as travelling, and the other way around.
  • hand and foot side: where the clip uses the other hand or foot than the caption names, the side was fixed.
• Floating and gait: walk, jog and run clips (no jump-like words) where the lowest foot stays more than 10 cm up for a whole second were left out (19 clips).

In this set 4,050 clips have a caption that differs from their prompt.

Caption style

Captions come from three sources (see Prompts) and read very differently. The template and seed captions are plain descriptions of the body's movement. Many written captions were drafted by a language model and read like prose instead. Counted by keyword in the captions (rough: a keyword can also be meant plainly), of the 48,977 written clips:

kind	written clips	example
manner words ("expertly", "resolute", "deliberate")	about 6,200	"spins slowly, arms sweeping in a circle, each sweep deliberate, as if summoning celestial bodies"
objects and props the data does not hold (weapons, tools, walls, furniture)	about 6,200	"swings a heavy hammer with a single, resolute arm movement, embedding it into the fantasy stone wall"
hand and finger detail the skeleton cannot show (no fingers)	about 4,000	"manipulates a rifle's actions, checks the chamber, reloads with a steady, methodical motion"
invented scenes ("fantasy", "virtual", "imaginary")	about 670	"hops in place, feet lightly tapping the virtual ground, demonstrating quick, controlled footwork"

About 14,600 written clips have at least one of these. The motion follows the body movement such a caption describes, roughly: there is no object, no hand detail and no scene in the data. These captions are kept as the model was trained on them. To train on plain descriptions only, use the template and seed clips, or the written clips without these words.

Limitations

• Kimodo-SOMA-RP-v1.1 saw BONES-SEED. NVIDIA reports that some BONES-SEED test sequences leaked into v1.1's training set (fixed in v1.2). Do not use models trained on this set to report results on the BONES-SEED test splits. Kimodo's own scores on the Kimodo benchmark remain optimistic.
• Kimodo cannot do some actions at all. Most of these were caught by eye or by measure, not all.
• Long text clips tend to slow down at the end.
• Path clips (text+path) slide the feet more than text clips.
• Prompts like "swims in the air" float the body above the floor.
• One neutral body, no body shape variety. Synthetic motion: it inherits Kimodo's style and biases.
• English only. Paraphrases are machine-written.

Provenance

The motion clips were generated by the Licensor using NVIDIA Kimodo (nvidia/Kimodo-SOMA-RP-v1.1), used under the NVIDIA Open Model License, under which NVIDIA claims no ownership rights in outputs. This Dataset is not endorsed by NVIDIA. No part of the Kimodo model is included in the Dataset. https://huggingface.co/nvidia/Kimodo-SOMA-RP-v1.1 , https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/

Texts from other sources:

Source	Used for	License	Link
SEED Timeline Annotations (NVIDIA)	the seed prompts	CC BY 4.0	https://huggingface.co/datasets/nvidia/SEED-Timeline-Annotations
Granite-3.3-2B-Instruct (IBM)	drafting part of the written prompts	Apache-2.0	https://huggingface.co/ibm-granite/granite-3.3-2b-instruct
Qwen3-8B, Qwen3-4B-Instruct-2507 (Qwen team)	paraphrases (checked), persona prompts	Apache-2.0	https://huggingface.co/Qwen/Qwen3-8B , https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507

How to cite

If you use SprongMotion in research or other published work, please cite the dataset as follows:

@misc{endendijk2026sprongmotion100k,
  author       = {Endendijk, Seger},
  title        = {SprongMotion-100K: generated human motion clips with captions},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/SegerEnd/SprongMotion-100K}},
  note         = {Seger Endendijk (segerend.nl)}
}

Please also credit Kimodo and the SEED texts: "Motions generated with NVIDIA Kimodo-SOMA-RP-v1.1. Seed captions from NVIDIA SEED Timeline Annotations (CC BY 4.0)."