otterbox901/q-1.3

Q 1.3

Q 1.3 is the research XL foundation with the QEVD add-on. It is a private copy for cubit labs' own research. Don't make this repository public, share it, or add anyone to it. The training data's terms forbid that (below).

folder	package	what
research-xl-v1/	foundation, size XL	pretrain-research-xl-01 step 95,000: 114.7M parameters, 16 blocks × 768, 12 heads
qevd@1/	research add-on on research-xl-v1	qevd-01, 30,000 steps: LoRA rank 16 (alpha 32, 3.1M parameters) on the attention and MLP layers, plus two label heads

Both are goata.motion-model/1 packages as written by goata-package, with each weights.pt checked against the sha256 in its package.json. The only change from the bucket copies: qevd@1/package.json names its foundation as ../research-xl-v1 instead of a gs:// address, so the stack loads from this folder.

Each folder also has weights.safetensors, the same tensors as its weights.pt (checked equal; keys flattened with ., and the original nesting plus the .pt file's sha256 are kept in the safetensors metadata). The root model.safetensors.index.json points at both, so the Hub shows the stack's parameter count: 118,542,520 (114,678,550 foundation + 3,863,970 adapter and heads). load_package still reads weights.pt.

Load

From a checkout of the Goata repository, in ml/:

from huggingface_hub import snapshot_download
from goata.model.package import load_package

path = snapshot_download("otterbox901/q-1.3")
stack = load_package(f"{path}/qevd@1")   # XL foundation + QEVD adapter and heads
foundation = load_package(f"{path}/research-xl-v1")

Inputs are the same as Q 1.1's: windows of 128 frames, 23 canonical joints, 8 features, at 30, 60 or 120 Hz (goata/model/spec.py).

Heads (qevd@1)

• exercise: one of 237 classes from QEVD's labels (exercises and everyday actions).
• how: multi-label, 483 classes: how the exercise is done (form, named mistakes, pace).

On 8,192 held-out QEVD windows at step 30,000:

head	measure	adapted	frozen XL + head only
exercise	accuracy	0.915	0.701
exercise	top-5	0.982	0.883
exercise	balanced accuracy	0.829	0.509
how	mean average precision	0.651	0.512 (base rate 0.149)
how	first choice right	0.701	0.555 (at random 0.164)

Foundation (research-xl-v1)

322.8 hours: Kinetics-700 (227.0 h, 2D pose from YouTube video), AddBiomechanics (54.1), DualSystem (11.0), CMU mocap (9.3), Riglet (6.4), Kuopio (5.1), AIST++ (4.1), Health&Gait (3.4), REHAB24-6 (1.5), Fukuchi running (0.7).

3DPW test (12 SMPL joint centres, hip-centred, MediaPipe input; not the published protocol): MPJPE 136.9 mm against the input's 135.6, so it doesn't improve on its input there. With 2D input only: 132.4. Q 1.1 (L) scored 119.9 on the same test.

Terms

Research only, never in a product, never published:

• QEVD (the add-on's only data): Qualcomm Data License Agreement – Research Use. It allows internal non-profit research only, and covers every model trained on it. If Qualcomm ends the licence, the data and these models must be deleted, this repository included.
• AIST++: the AIST Dance DB terms forbid commercial use without written consent.
• REHAB24-6: CC BY-NC 4.0.
• Kinetics-700: annotations CC BY 4.0, the videos are YouTube's.