LibriBrain

The LibriBrain 2025 dataset family provides MEG-based speech and language tasks with download/caching support from Hugging Face.

Common Arguments

• data_path: local root where files are stored / downloaded

• preprocessing / preprocessing_str: expected preprocessing string in filenames

• standardize: z-score channels using per-run stats

• include_run_keys: list of run keys to include (see constants.RUN_KEYS)

• include_info: include an info dict in each sample

• download: if True (default), fetch missing files via Hugging Face

Task-based entry point

from pnpl.datasets import LibriBrain
from pnpl.tasks import SpeechDetection

ds = LibriBrain(
    data_path="./data/LibriBrain",
    task=SpeechDetection(tmin=0.0, tmax=0.2),
    partition="train",
    include_info=True,
)

print(len(ds))

The task object controls sample collection and label semantics. Public task classes live in pnpl.tasks.

Wrapper datasets

Speech (binary time series)

from pnpl.datasets import LibriBrainSpeech
from pnpl.datasets.libribrain2025 import constants

ds = LibriBrainSpeech(
    data_path="./data/LibriBrain",
    preprocessing_str="bads+headpos+sss+notch+bp+ds",
    include_run_keys=[constants.RUN_KEYS[0]],
    tmin=0.0,
    tmax=0.2,
    include_info=True,
)

print(len(ds))

Each item returns (data: float32[channels,time], labels: int[time], info: dict) when include_info=True.

Phoneme (classification)

from pnpl.datasets import LibriBrainPhoneme
from pnpl.datasets.libribrain2025 import constants

ds = LibriBrainPhoneme(
    data_path="./data/LibriBrain",
    preprocessing_str="bads+headpos+sss+notch+bp+ds",
    include_run_keys=[constants.RUN_KEYS[0]],
    tmin=-0.2,
    tmax=0.6,
)
print(len(ds))

Each item returns (data: float32[channels,time], label_id: int64).

LibriBrainWord and LibriBrainSentence are also available as dataset-specific wrappers.