The 2026 PNPL Competition focuses on a single, ambitious task: Word Classification. Given a segment of MEG brain data recorded while a subject listens to spoken English, predict the specific word being heard from a fixed 50-word vocabulary (the task is not open-vocabulary).
Evaluation metric
Submissions are scored by top-10 balanced accuracy (BAcc@10) over the 50-word vocabulary — a prediction counts as correct if the true word is among your model's ten most likely candidates, balanced across classes so that rarer words matter as much as common ones.
Final rankings are determined on independent holdout data, recorded with stimuli entirely different from anything released to participants.
Two tracks: Deep & Broad
Both tracks share the same word-classification task, but ask different questions about how neural speech decoders generalise. You may enter either or both.
Deep
Word classification on a single, deeply-sampled subject (subj0). How far can you push accuracy when data from one brain is abundant?
Training data: extensive recordings from subj0 spanning audiobooks, phonetically balanced speech corpora (TIMIT, MOCHA-TIMIT), and narrative podcasts.
Broad
Word classification across 32 held-out subjects (subj1–subj32). How well can a model generalise to new brains from limited data?
Training data: varying amounts of labelled data per subject, grouped by how much fine-tuning data is available (see below).
View on Kaggle →Track 2 data groups
To study how performance scales with the amount of fine-tuning data, the 32 Broad subjects are split into three groups:
| Group | Subjects | Labelled data per subject |
|---|---|---|
| 100% | 12 | ~40 min of audiobook listening |
| 50% | 10 | ~20 min of audiobook listening |
| 25% | 10 | ~10 min of audiobook listening |
The Dataset: LibriBrain100
This year's competition uses LibriBrain100, a major expansion of the original LibriBrain dataset to over 100 hours of MEG data. It pairs the deeply-sampledsubj0 recordings behind Track 1 with data from 32 additional subjects (subj1–subj32) behind Track 2, all listening to naturalistic spoken English. Together they keep the depth of the original LibriBrain while adding breadth across 33 subjects in total.