README.md

---
language:
- en
license: cc-by-nc-4.0
task_categories:
- other
---

# LibriBrain (Sherlock Holmes 1–7)

[Paper](https://huggingface.co/papers/2506.02098) | [Code](https://github.com/neural-processing-lab/pnpl)

This repository contains the LibriBrain data organised by book: MEG recordings (`.h5`), event annotations (`.tsv`), and the audiobook stimulus audio (`.wav`). 

LibriBrain was first open-sourced as part of the [2025 PNPL Competition](https://libribrain.com/).

In addition, LibriBrain is used as a fine-tuning dataset in the paper ["MEG-XL: Data-Efficient Brain-to-Text via Long-Context Pre-Training"](https://huggingface.co/papers/2602.02494) to evaluate word decoding from brain data.

## Sample Usage
The easiest way to get started with the dataset is using the [pnpl Python library](https://github.com/neural-processing-lab/pnpl). There, the following two datasets are available:

### LibriBrainSpeech
This wraps the LibriBrain dataset for use in speech detection problems.
```python
from pnpl.datasets import LibriBrainSpeech

speech_example_data = LibriBrainSpeech(
    data_path="./data/",
    partition="train"
)

sample_data, label = speech_example_data[0]

# Print out some basic info about the sample
print("Sample data shape:", sample_data.shape)
print("Label shape:", label.shape)
```

### LibriBrainPhoneme
This wraps the LibriBrain dataset for use in phoneme classification problems.
```python
from pnpl.datasets import LibriBrainPhoneme

phoneme_example_data = LibriBrainPhoneme(
    data_path="./data/",
    partition="train"
)
sample_data, label = phoneme_example_data[0]

# Print out some basic info about the sample
print("Sample data shape:", sample_data.shape)
print("Label shape:", label.shape)
```

### Usage in MEG-XL
To fine-tune the MEG-XL model on the LibriBrain dataset for word decoding, you can use the following command from the [official repository](https://github.com/neural-processing-lab/MEG-XL):

```bash
python -m brainstorm.evaluate_criss_cross_word_classification \
  --config-name=eval_criss_cross_word_classification_libribrain \
  model.criss_cross_checkpoint=/path/to/your/checkpoint.ckpt
```

Note: For the MEG-XL repo, you will need to adjust the dataset paths in the configuration files to point to your local download of the data. The pnpl library includes automatic downloads from HuggingFace.


## Repository structure

Data are organised into seven top-level directories:

- `Sherlock1/`
- `Sherlock2/`
- …
- `Sherlock7/`

Each `Sherlock{i}` directory contains:

- `Sherlock{i}/derivatives/events/` — event annotation files (`.tsv`)
- `Sherlock{i}/derivatives/serialised/` — MEG recordings (`.h5`)
- `Sherlock{i}/stimuli/audio/` — stimulus audio (`.wav`)

## Stimulus audio (LibriVox)

The spoken-audio stimuli are derived from **LibriVox** public-domain recordings of the first seven Sherlock Holmes books (recording versions linked below). The stimuli are provided in this repository as WAV files converted from the LibriVox downloads.

### LibriVox source URLs (recording versions)

1. https://librivox.org/a-study-in-scarlet-version-6-by-sir-arthur-conan-doyle/
2. https://librivox.org/the-sign-of-the-four-version-3-by-sir-arthur-conan-doyle/
3. https://librivox.org/the-adventures-of-sherlock-holmes-version-4-by-sir-arthur-conan-doyle/
4. https://librivox.org/the-memoirs-of-sherlock-holmes-by-sir-arthur-conan-doyle-2/
5. https://librivox.org/the-hound-of-the-baskervilles-version-4-by-sir-arthur-conan-doyle/
6. https://librivox.org/the-return-of-sherlock-holmes-by-sir-arthur-conan-doyle-2/
7. https://librivox.org/the-valley-of-fear-version-3-by-sir-arthur-conan-doyle/

### Audio format

The WAV files in this repository are:
- WAV (PCM), mono (1 channel), 22,050 Hz, 16-bit signed integer PCM

Example conversion command (SoX):

```bash
sox "INPUT_FROM_LIBRIVOX.mp3" -c 1 -r 22050 -b 16 "OUTPUT.wav"
```

### Citation

If you use this dataset, please cite the LibriBrain paper:

```bibtex
@article{ozdogan2025libribrain,
  author    = {Özdogan, Miran and Landau, Gilad and Elvers, Gereon and Jayalath, Dulhan and Somaiya, Pratik and Mantegna, Francesco and Woolrich, Mark and Parker Jones, Oiwi},
  title     = {{LibriBrain}: Over 50 Hours of Within-Subject {MEG} to Improve Speech Decoding Methods at Scale},
  year      = {2025},
  journal   = {NeurIPS, Datasets \& Benchmarks Track},
  url       = {https://arxiv.org/abs/2506.02098},
}
```

If you use this data with the MEG-XL framework, please also cite:

```bibtex
@article{jayalath2026megxl,
  title={{MEG-XL}: Data-Efficient Brain-to-Text via Long-Context Pre-Training},
  author={Jayalath, Dulhan and Jones, Oiwi Parker},
  journal={arXiv preprint arXiv:2602.02494},
  year={2026}
}
```