MEG Speech Decoding

Magnetoencephalography (MEG) records the magnetic fields produced by neural activity with millisecond resolution, and decoding heard speech from those recordings is a frontier problem in neuro-AI. The signal is faint and noisy, hundreds of sensor channels mix overlapping neural sources, and word-evoked responses vary across recordings — so a decoder that memorizes session quirks collapses on held-out data.

This task uses continuous LibriBrain recordings aligned to audiobook speech: given a recording and the sample index where a word was heard, the decoder must rank the vocabulary by how likely each word is to be the one the listener heard. The design space is deliberately open — from classical feature pipelines to end-to-end networks or adapting pretrained speech and MEG foundation models — but the result must survive sealed evaluation slices that stress rare words and long-duration generalization.

Build a word decoder and keep the complete replayable deliverable under /app/meg_decoder/: predict.py, model.py, run_summary.json, and a trained checkpoint/.

• predict.py is invoked with --data-dir, --checkpoint, and --output-path, reads MEG recordings, event metadata, sensor metadata, and the vocabulary, and writes one JSONL row per input example_id.
• Each row must contain a deterministic ranked list of at least ten unique vocabulary IDs (word_ids); duplicate or out-of-vocabulary IDs invalidate the row.
• A full inference batch must finish within 30 minutes without internet or external services.
• Rankings must materially respond to the supplied MEG signals rather than relying on event identifiers or fixed vocabulary priors — a counterfactual probe enforces this.
• The complete /app/meg_decoder/ directory must be no larger than 2.5 GB.

Labeled LibriBrain development data under /data: training and validation splits with continuous MEG arrays in recordings.zarr, word-onset events with labels in events.parquet, sensor metadata, the complete vocabulary.json, and — for the training split only — the aligned stimulus audio. Optional pretrained checkpoints are staged offline under /models: the facebook/w2v-bert-2.0 speech encoder, the pnpl/MEG-XL MEG foundation model, and google-t5/t5-large. The pinned Python stack includes torch and torchaudio 2.6.0 (CUDA), transformers, mne, librosa, zarr, pandas, and scikit-learn, with one GPU available. A fail-closed scaffold sits at /app/meg_decoder/ and /app/validate_predictions.py checks outputs locally. The machine is offline; a sandbox-timer CLI reports the remaining budget.

A clean-room verifier evaluates a read-only snapshot of /app/meg_decoder/ on sealed hidden data, stripping every label-bearing column from the events it hands to the decoder and locking /data, /models, and the workspace to root during replay. Inference runs as the unprivileged agent user with a 30-minute cap on the full batch.

• Quality is scored per workload — held-out recordings, recording shift, rare words, and long duration (weighted 0.4/0.2/0.2/0.2) — using a chance-adjusted composite of macro top-10 accuracy (60%), macro MRR (25%), and macro top-1 accuracy (15%), combined across workloads by a smoothed geometric mean and calibrated onto the final reward scale.
• A determinism probe reruns the full batch and requires identical top-10 rankings.
• A signal-counterfactual probe time-shifts the real MEG while preserving event IDs and onsets; at least 50% of each top-10 ranking must change, defeating decoders that ignore the signal.
• Contract and safeguard gates are binary: format violations, missing or duplicate predictions, a failed or nondeterministic replay, insufficient signal turnover, or an oversized deliverable zero the result.