On 25 June 2026 Euan Ashley’s group at Stanford published OpenMHC: Accelerating the Science of Wearable Foundation Models on arXiv, a dataset and benchmark for wearable foundation models, reaching its third version on 8 August. The abstract announces “open-source code and model weights” and data under broad access. This is the state of the repositories on 12 August.

What it is

OpenMHC, in full OpenMyHeartCounts, derives from more than ten years of collection through the My Heart Counts study app. The dataset as stated in the paper:

  • 67 million hours of wearable signal
  • 19 sensor channels — steps, heart rate, sleep, workouts
  • up to 169 linked variables — health, lifestyle, mood, behaviour
  • 11,894 participants of the study’s 16,993, those who consented to broad sharing

Alongside the data there are implementations of recent foundation models, among them adaptations of Google’s LSM2 and Apple’s WBM, and a unified benchmark on three tracks: prediction of health and behaviour variables, imputation of missing data and time-series forecasting. Repository github.com/AshleyLab/OpenMHC, leaderboard on a Hugging Face Space.

Release status

The README carries a release plan as a set of boxes. On 12 August four are ticked:

  • paper on arXiv
  • evaluation code
  • checkpoints on Hugging Face
  • benchmark Space

Four are not:

  • dataset — “will be made available to all qualified researchers free of charge upon acceptance of the paper”
  • cleaned-up training infrastructure — estimate “July-August”
  • usage examples for the pre-trained models — estimate “July-August”
  • adaptor for Apple HealthKit exports — estimate “July-September”

Code

The main repository github.com/AshleyLab/OpenMHC was created on 28 April 2026, received its last push on 9 August and has 39 stars. It holds the evaluation code, the API for running your own methods on the benchmark and the reference implementations.

The code licence is declared twice and never granted. pyproject.toml at line 11 carries license = { text = "MIT" } and the README’s ## License section repeats “Code: MIT”. There is no licence file: LICENSE, LICENSE.md, LICENSE.txt and COPYING all answer 404, and GitHub’s detector reports no licence. MIT requires its own text and the copyright notice to travel with copies, so the declaration alone does not put anyone in a position to reuse the code.

Part of the training code is here: src/imputation_training/ and src/forecasting_training/, the mhc-impute-train and mhc-forecast-train commands declared in [project.scripts], the Hydra configurations under configs/training/ and configs/forecasting_train/ and the SLURM scripts in jobs/. The unticked box concerns the cleaned-up version, and for the rest the README points to a second repository: “The research-grade codebased can be found here https://github.com/NarayanSchuetz/OpenMHC (particularly relevant for training infra, until we ported that properly)”. That repository is under the MIT licence, has 4 stars and was created on 7 May 2026, the date of its only push. The published checkpoints are from late May and June.

Weights

14 checkpoints published under the MyHeartCounts organisation on Hugging Face between 29 May and 22 June 2026. The format is not uniform: 5 repositories publish a PyTorch .ckpt checkpoint, 8 a PyPOTS .pypots artefact and one, openmhc-chronos2-fc, a model.safetensors under checkpoint/. openmhc_manifest.json is in all 14; normalization_stats.json in 9, while openmhc-dlinear-fc, openmhc-mixlinear-fc and openmhc-segrnn-fc use standard_scaler_stats.json with training_config.json, and openmhc-chronos2-fc and openmhc-toto-fc publish no normalisation statistics.

The licences declared in the model cards are not uniform:

  • OpenRAIL — 8 checkpoints: openmhc-lsm2-daily, openmhc-lsm2-weekly, openmhc-lsm2-weekly-sparse, openmhc-brits-imp, openmhc-dlinear-imp, openmhc-dlinear-7day-imp, openmhc-fedformer-imp, openmhc-timesnet-imp
  • CC BY 4.0 — 6 checkpoints: openmhc-chronos2-fc, openmhc-dlinear-fc, openmhc-mixlinear-fc, openmhc-segrnn-fc, openmhc-toto-fc, openmhc-wbm-dp

The split follows the tracks: OpenRAIL covers the 8 checkpoints of the imputation track, CC BY 4.0 the 5 forecasting ones and the single downstream prediction one. The README’s ## License section, however, states “Models OpenRAIL” without distinction, and 6 checkpoints out of 14 sit under a different licence: anyone reusing the weights has to read each model card.

Data

DATASET.md gives Harvard Dataverse as the dataset’s home: about 38 GB, with the DOI listed as “Available upon paper acceptance”. The development subset version="xs" downloads today, 593 users and about 1.9 GB, that is 5.0% of the cohort, meant for quickstarts and smoke tests. The API exposes openmhc.download_dataset(version=
, dest=
) and writes into the destination a dataset_version.json marker that the evaluation checks back against.

Access to the full release is regulated. The paper reports Stanford’s ethical approval, protocol IRB-31409, and a digital consent in which each participant chose between “narrow” sharing, with Stanford alone, and “broad”, with qualified researchers worldwide. The released dataset is made of the “broad” group alone: 11,894 users and 67 million hours out of 16,993 users and more than 80 million hours collected. Where distribution is bound to a data use agreement, access requires a Dataverse API token.

No OSI licence applies to individual health data. The constraint comes from the consent the participants signed and dictates the shape of the distribution, the same one I described designing clinical RAG systems on-premise.

Against the Open Source AI Definition

The Open Source AI Definition 1.0 requires three components together. Code “shall be made available under OSI-approved licenses”; for data information and parameters the formula is wider, “under OSI-approved terms”. The state on 12 August:

  • Data information — the requirement concerns the description, and expressly contemplates unshareable data, of which it asks “the complete description of all data used for training, including (if used) of unshareable data, disclosing the provenance of the data, its scope and characteristics”. The paper describes cohort, channels, variables, 92 device models and a thirteen-year span; the description is not published under OSI-approved terms.
  • Code — “the complete source code used to train and run the system”. Evaluation and training are both published, with no licence file in either of the repositories that hold them.
  • Parameters — published, under OpenRAIL for 8 checkpoints and CC BY 4.0 for 6.

The two licences miss the mark for different reasons. OpenRAIL carries behavioural use restrictions, which fall under clause 6 of the Open Source Definition — “No Discrimination Against Fields of Endeavor” — and among the recurring reasons for rejection OSI lists ethical and non-commercial clauses. CC BY 4.0 does not restrict use and raises no problem on that front, but it is a licence for creative works: it grants no patent rights and is not designed for software artefacts.

It is the same boundary measured on DeepSeek-R1 and on GPT-NeoX, crossed in full by TinyLlama.

On the parts that do not depend on participant consent, three things remain to be done: a LICENSE file in the code repositories, one licence across the 14 checkpoints matching what the README declares, and a date for the dataset independent of the review outcome.

Limits

The above describes the state of the repositories on one day, and the unticked boxes carry estimates that run out in September: the same check a month from now may give a different result. I did not run the benchmark or reproduce the results, which the 5.0% subset does not allow; on model performance this article says nothing. The licence classification comes from the fields declared in the model cards and the README, not from legal advice, and whether a release conforms to the OSI definition is for OSI to determine, not for a reader with the APIs.

Cover image: Étienne-Jules Marey’s direct sphygmograph, from La mĂ©thode graphique, Paris 1878, p. 281 — public domain — https://commons.wikimedia.org/wiki/File:Marey_Sphygmograph.jpg