novogaia/massive-v1-ms2-100m-stratified-x16

MassIVE v1 MS2 100M Stratified HDF5 Shards

This dataset contains a 100,000,000-row chunk-aligned stratified sample from the MassIVE v1 MS2 HDF5 handoff file.

• Source object: gs://metal-repeater-411410-spectra-checkpoints/test_data_massive/v1_handoff/MassIVE_v1_ms2.hdf5
• Source rows: 1,620,804,459
• Sampled rows: 100,000,000
• Shards: 16 HDF5 files, 6,250,000 rows per shard
• Compression: HDF5 gzip level 1
• Sampling unit: source HDF5 row chunk, 256 rows per source chunk
• Stratification dataset: precursor_mz
• Stratification bins: precursor m/z upper edges 100, 200, ..., 1000

The shard set includes all 33 datasets from the source file. Row-aligned datasets are sampled to 100M rows, while non-row metadata datasets under metadata/ are copied unchanged into each shard.

Files

• fdataloader_shards.json: shard manifest and sampling metadata
• shard_00000.hdf5 ... shard_00015.hdf5: HDF5 shard files

Row-Aligned Datasets

MS level, RT, acquisition_type, activation_energy, charge, collision_energy, dformat, file_id, file_name, global_group_id, group_id, massive_id, positive polarity, precursor intensity, precursor scan number, precursor target intensity, precursor_mz, precursor_target_mz, scan_number, spectra_id, spectrum, type, window lo, window uo.

Static Metadata Datasets

metadata/#TBXICs, metadata/#TBXICs(1), metadata/MSLevelOrder, metadata/Ordered RT, metadata/TBXICs mean stdev, metadata/TBXICs median stdev, metadata/file_name, metadata/instrument name, metadata/name.

Sampling Note

The sample is chunk-aligned rather than row-random. The sharder first scans precursor_mz, assigns source chunks to precursor bins, samples chunks across bins, then copies the selected source rows into the output shards. This avoids decompressing nearly every gzip-compressed source spectrum chunk while preserving precursor-m/z diversity at chunk level.

Loading

The shards are directly readable with HDF5 tools. They are also compatible with fdataloader:

from fdataloader import MassiveHdf5ShardDataset, build_dataloader

dataset = MassiveHdf5ShardDataset("fdataloader_shards.json")
loader = build_dataloader(dataset, batch_size=256, num_workers=8, shuffle=True)

For distributed training, pass world_size and rank, or launch with standard distributed environment variables.