How to access the IBL Brain-Wide Map LFP dataset

The Brain-Wide Map (BWM) LFP dataset holds 699 recordings of 384 channels in one .h5 file, published in three tiers that differ only in the wavelet threshold α (choosing α). Every file has the same layout and the same 699 recordings.

Tier	α	File	Size	File / float32 at 250 Hz
small	14	`lf_compressed_v04_a14_small_all_bwm.h5`	7.7 GB	0.57 %
default	7	`lf_compressed_v04_a07_default_all_bwm.h5`	14.9 GB	1.10 %
fine	2.5	`lf_compressed_v04_a2p5_fine_all_bwm.h5`	31.9 GB	2.36 %

small decodes wheel speed, choice, feedback and brain region about as well as the retired v03 mild tier at 0.38× its size; fine recovers the scores of the uncompressed signal for behaviour. Start with default. The v03 files (lf_compressed_all_bwm.h5, lf_compressed_mild_all_bwm.h5, lf_compressed_aggressive_all_bwm.h5) are format 1 and superseded.

Browse online

lfpack.internationalbrainlab.org hosts a live web viewer over the dataset: pick a recording (searchable by pid) or filter by brain region (acronym or name, e.g. VISp or Isocortex, matching the region and its Allen descendants), then pan, zoom and adjust gain. No download needed.

Download

No AWS credentials are needed; the files are in the public IBL bucket:

from one.remote.aws import s3_download_file

s3_download_file(
    source="resources/ibl-agent-data/lf_compressed_v04_a07_default_all_bwm.h5",
    destination="lf_compressed_v04_a07_default_all_bwm.h5",
)

s3_download_file shows a progress bar and skips the download if the local file already has the right size.

Limitations

Saturation. ADC-clipped stretches are flagged and muted in the output. 446 of the 699 recordings have at least one; check your window before trusting amplitude.
High frequencies. Power above roughly 20–30 Hz is traded for compression, more so at small than at fine.
CSD. The archives were denoised with the v04 Cadzow, whose CSD has seams every 16 channels, and the codec smooths fine spatial structure: check CSD features that persist over a chunk against uncompressed data. Cadzow in lfpack ≥ 1.1 has no seams; no BWM archive uses it yet.
Low frequencies. Per-channel impedance and analog-filter differences leave channel-to-channel amplitude outliers below 1.5 Hz that the pre-processing does not fully correct.

Saturation across the 699 recordings: saturated fraction per recording (left) and the 15 most affected (right).

Before you start

Align with trial events via sr.times. Each recording stores a sync-corrected session clock; do not use np.arange / sr.fs. See Read lfpack files.
Bin channels for brain-wide sweeps. bin_channels=4 reduces each recording to 96 spatial bins. See binned reads.

from lfpack import LFPackReader

file_lfpack = "lf_compressed_v04_a07_default_all_bwm.h5"
pid = "dab512bd-a02d-4c1f-8dbc-9155a163efc0"

sr = LFPackReader(file_lfpack, recording=pid)
traces = sr[: int(10 * sr.fs)]     # (2500, 384) float32, volts
t = sr.times[: len(traces)]        # session-clock seconds

LFPackReader.recordings(file_lfpack) lists the 699 pids.

Quality control: saturation

Each recording carries a table of saturated spans, detected on the raw band before any filtering and zeroed in the decompressed output:

sr.saturation             # DataFrame: start_sample, stop_sample (raw-rate samples)
sr.saturation_summary     # saturated_fraction, n_intervals, total_saturated_sec, muted, ...
sr.saturation_times()     # the table plus start_time, stop_time (session clock)

n = int(10 * sr.fs)
mask = sr.saturation_mask[:n]       # (n,) bool at the reader's rate, True where saturated
traces_clean = sr[:n][~mask]

All four return empty or default values for recordings without saturation detection, so the same code runs on the whole dataset.

Channel brain locations

Brain locations are embedded in each file; no network call is needed:

ch = sr.channels
ch["acronym"][:5]    # Allen acronym per channel
ch["atlas_id"][:5]   # Allen CCF structure IDs, int32
ch["labels"][:5]     # 0 good, 1 dead, 2 noisy, 3 outside the brain
ch["x"][:5], ch["y"][:5], ch["z"][:5]   # ML, AP, DV in metres

With bin_channels, coordinates are averaged and regions take the within-group mode; sr.channels_full always returns the raw electrodes.

Visualise with viewephys

uv pip install "viewephys[lfpack]"
viewephys -f lf_compressed_v04_a07_default_all_bwm.h5

The viewer has a searchable recording selector, colours the depth axis by brain region from the embedded annotations, and adds a CSD view.

viewephys density view showing 384-channel LFP data with a brain region colour bar and a searchable recording selector — viewephys on a BWM file: brain regions colour the depth axis and the selector switches between the 699 recordings.