docs/upload_to_huggingface.md

Upload to Hugging Face

This folder is the structured, no-Miqus upload tree. It is generated from the cleaned staging data via hard-links (so re-uploading is cheap and does not duplicate ~86 GB on disk).

One-shot upload

hf upload-large-folder <hf-user-or-org>/PULSE \
    /path/to/hf_upload_pulse_structured_nomiqus \
    --repo-type dataset \
    --num-workers 8 \
    --exclude '.cache/**' \
    --exclude '*Miqus*.avi'

LFS rules in .gitattributes cover the large CSVs (IMU, MoCap, raw Qualisys TSV), the auxiliary MP4s, and all standard binary formats. Do not delete .gitattributes before upload — without it the large CSVs would be pushed as plain git blobs and the repo would exceed the per-file limit.

Migrating from the old top-level v*/ layout

The previous upload had volunteer folders directly at the root (<repo>/v1/s1/...). This release moves them under data/ (<repo>/data/v1/s1/...).

hf upload-large-folder does not delete files that are no longer present in the local tree. If the same Hugging Face repo already contains the old layout, you must clean up the old paths explicitly so the dataset card globs and Croissant includes resolve unambiguously:

# List old top-level v*/ paths still on the remote
hf api repo-files <hf-user-or-org>/PULSE --repo-type dataset \
    | grep -E '^v[0-9]+/'

# Remove them (irrevocable; review the list first)
hf repo-files delete <hf-user-or-org>/PULSE \
    --repo-type dataset \
    'v*/s*/*' 'batch_alignment_summary.json' 'modality_coverage.xlsx' \
    'annotations/_run_summary.json' 'annotations/segment_counts.*' \
    'annotations/logs/*'

Note that the raw Qualisys TSV has also moved one level deeper, from data/v*/s*/aligned_v*_s_Q.tsv (in the previous staging tree) to data/v*/s*/raw/aligned_v*_s_Q.tsv here. If you re-uploaded an interim copy that still had the TSV at the old location, clean those up too:

hf repo-files delete <hf-user-or-org>/PULSE \
    --repo-type dataset 'data/v*/s*/aligned_v*_s_Q.tsv'

Verifying the upload

After upload completes, sanity-check that:

1. huggingface.co/datasets/<repo>/tree/main shows the four top-level dirs (annotations/, data/, docs/, metadata/) plus the licenses, README, and croissant.json — and no stray v*/ at the root.
2. The dataset viewer can load each per-modality config:

   from datasets import load_dataset
   for cfg in ["emg", "imu", "mocap", "eyetrack", "pressure"]:
       ds = load_dataset("<repo>", cfg, split="train", streaming=True)
       print(cfg, next(iter(ds)).keys())
   

3. Large files were stored via LFS:

   hf api file-info <repo> data/v1/s1/aligned_imu_100hz.csv \
       --repo-type dataset
   # Should report `lfs: {...}` rather than a raw blob sha.