nm000104 NEMAR-native dataset

emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography

emg2qwerty is the largest public surface electromyography (sEMG) dataset to date, comprising 1,135 sessions from 108 participants performing touch typing on a QWERTY keyboard. The dataset captures wrist-based sEMG signals (32 channels, 2000 Hz sampling rate) synchronized with keystroke ground truth, totaling 346.4 hours of data and 5.26 million keystrokes. Designed to enable keyboard-free text input through decoding of typing intent from neuromuscular activity, the dataset supports research in sequence-to-sequence learning, cross-user generalization, domain adaptation, and neuromotor interfaces for AR/VR and accessibility applications.

EMG Zarr unverifiable
Issues GitHub

Download this dataset

Archive (>100 GB) removed to save space; use the per-file direct download (#752).. Use one of the streaming methods below — all resumable. Full download guide →

  1. NEMAR CLI recommended

    Pulls the pinned version + annexed data and resumes cleanly. Install nemar-cli → 

    nemar dataset download nm000104
  2. DataLad

    Clone the dataset repo and fetch file content on demand. Docs → 

    datalad clone https://github.com/nemarDatasets/nm000104 nm000104
cd nm000104 && datalad get .
  3. git-annex

    Plain git + git-annex against the dataset repo. Docs → 

    git clone https://github.com/nemarDatasets/nm000104 nm000104
cd nm000104 && git annex get .
  4. Direct files (wget / curl / rclone)

    Every file with a stable, range-resumable URL from the manifest. Needs curl, jq, wget (or rclone/aria2c). Docs → 

    curl -s https://data.nemar.org/nm000104/v2.0.0/manifest.json | jq -r '.[].bytes_url' > urls.txt
wget -xc -i urls.txt

Compute on this dataset

Two routes today, with a third (in-browser one-click submission) landing soon.

  1. NeuroScience Gateway (NSG) portal.

    NSG runs EEGLAB / Brainstorm / MNE pipelines on supercomputing time donated by SDSC. Create an account, point a job at this dataset's S3 prefix (s3://nemar/nm000104), and submit.
    nsgportal.org →

  2. Local processing with nemar-cli.

    Pull the dataset to your machine and run any toolbox locally. Honors the published version pinning. 

    npm install -g nemar-cli
nemar dataset clone nm000104
cd nm000104 && nemar dataset get
  3. Just the files.

    rclone, aria2c, or any HTTPS client works against data.nemar.org/nm000104/ — the manifest carries presigned S3 URLs.

Direct compute access is coming soon. One-click NSG submission from this page is scoped for a follow-up phase. Tracked on nemarOrg/website#6.

Citations

    Loading demographics…

    Files

    Loading file index…

    Signal viewer
    How to use the data (for agentic research) license, citation, download commands, Zarr access

    What it is

    Modalities
    EMG
    Participants
    108
    Size
    223 GB
    Tasks
    typing

    License and terms

    License
    CC-BY-NC-SA-4.0
    Note
    Non-commercial use only (CC-BY-NC-SA-4.0).
    Recommended citation
    Sivakumar, V., Seely, J., Du, A., Bittner, S. R., Berenzweig, A., Bolarinwa, A., Gramfort, A., & Mandel, M. I. (2026). emg2qwerty: A Large Dataset with Baselines for Touch Typing using Surface Electromyography (Version v2.0.0) [Data set]. NEMAR. https://doi.org/10.82901/nemar.nm000104

    Where the bytes are

    Latest version (always current)
    https://data.nemar.org/nm000104/latest/
    This version (v2.0.0)
    https://data.nemar.org/nm000104/v2.0.0/

    How to download

    The dataset
    nemar dataset download nm000104 Clones and fetches in one step. Content under stimuli/ and derivatives/ is skipped by default because those trees can be large; add --stimuli --derivatives for the whole thing.
    A subset, one step
    nemar dataset download nm000104 --subjects sub-01,02 Also filters by --sessions, --tasks, --runs, --datatypes, --include and --exclude.
    A subset, step 1
    nemar dataset clone nm000104 Clones git-annex pointers only; fetches no file content. Creates ./nm000104.
    A subset, step 2
    cd nm000104 The get command below reads the clone's annex, so it only works from inside the clone.
    A subset, step 3
    nemar dataset get <files> Pulls the files you actually need. Skips stimuli/ and derivatives/ unless the path you ask for is under one of them.
    One small file
    https://data.nemar.org/nm000104/v2.0.0/participants.tsv A direct HTTPS fetch works for any single file.

    Assess fit without downloading

    Participants table
    https://data.nemar.org/nm000104/v2.0.0/participants.tsv
    Dataset description
    https://data.nemar.org/nm000104/v2.0.0/dataset_description.json
    Directory listing
    https://data.nemar.org/nm000104/v2.0.0/?format=json
    Catalog record
    https://api.nemar.org/datasets/nm000104

    Working with the Zarr copy

    1. Start at the index
    https://zarr.nemar.org/nm000104/zarr/index.json The mandatory entry point. Never hardcode a bucket path.
    2. Pick a store entry
    stores[].zarr, stores[].groups[].name These two fields exist in every index format version, so a recipe that keys on them works against the whole catalog while the back conversion is still in flight.
    3. Build the store URI
    s3://nemar/nm000104/zarr/{store.zarr} Derivable from the store entry alone. An index at format_version 3 or later also publishes contract_base, data_base and s3_uri; use them when they are there, never require them.
    4. Open the store anonymously
    zarr.open_group(store=..., mode="r", zarr_format=3) Anonymous FsspecStore.from_url in region us-east-2, no credentials. zarr_format=3 is required: without it zarr-python probes for Zarr v2 sidecars, and because anonymous ListBucket is denied, S3 answers a missing key with 403 rather than 404 and the open raises.
    5. Read the level-0 array
    root[store.groups[0].name]["0"] Level 0 is the full-rate signal. Never read a view/ array for inference; those exist for display.
    6. Dequantize the samples
    physical = digital * scale + offset scale and offset are attributes of the level-0 array, one entry per channel; the unit is on the group's channels attribute.
    7. Slice, don't download
    signal[0:4, 0:500] Stream a window of channels and samples; download only when you will touch most of the array.
    8. Know the HTTP contract
    index.json Only index.json is always proxied and edge-cached. A plain GET for a store object, manifest.json or events.parquet 302s to the public S3 object for non-browser clients, so follow redirects, and HEAD is never redirected.
    9. Read the attribution before reuse
    root.attrs["nemar"] The store carries its own dataset id, DOI, license, citation and source commit.
    10. Filter for pipelines
    has_zarr=1 This is the converted filter. has_zarr_verified is the stricter one, and its result set can be empty until the daily fidelity sweep reaches a dataset; verification is reported, never a precondition for serving (nemar-cli ADR 0005).