{ "schema_version": "bci-report-neurolm-eegmat-update-v1", "release_id": "neurolm-eegmat-update-20261020", "generated_at": "2026-10-20", "status": "aggregate_preview", "title": "NeuroLM’s frozen features on arithmetic and rest", "question": "Do NeuroLM’s frozen EEG features separate rest from mental arithmetic better than band power?", "display_track": "Frozen-feature development pilot", "placement": { "page": "/topics/pilot-studies/", "anchor": "neurolm-eegmat", "pointer_from": "/protocols/arithmetic-rest/", "counted_in_pilots": false, "leaderboard": false, "core_matrix_row": false, "new_dataset_count": 0, "new_model_count": 0, "text": "A separate panel on the pilot studies page, outside the pilot count and every leaderboard and model ranking; the core Arithmetic & rest matrix keeps its rows." }, "metric_units": "balanced accuracy, macro-F1 and chance levels are proportions in [0,1], and paired differences are differences of proportions: multiply by 100 for percent or percentage points; dimensions and counts of people, recordings, windows, channels, folds, fits and resamples are whole numbers; null means not computed, never zero", "scope": "NeuroLM-B stays frozen. One fixed ridge head (alpha 100) is fitted on each of three representations of the same 2,160 two-second windows from 36 people, in five participant-held-out folds: 15 fits and no tuning. Scores are equal-person means; the paired differences carry participant-bootstrap intervals.", "cohort": { "dataset": "EEGMAT", "name": "EEG During Mental Arithmetic Tasks", "version": "1.0.0", "source": "https://physionet.org/content/eegmat/1.0.0/", "people": 36, "recordings": 72, "conditions": 2, "retained_seconds": 60, "window_seconds": 2, "windows_per_recording": 30, "windows": 2160, "channels": 19, "channel_names": [ "FP1", "FP2", "F3", "F4", "F7", "F8", "T3", "T4", "C3", "C4", "T5", "T6", "P3", "P4", "O1", "O2", "FZ", "CZ", "PZ" ], "source_sampling_hz": 500, "sampling_hz": 200, "folds": 5, "fold_sizes_people": [ 8, 7, 7, 7, 7 ], "split": "whole participants held out: participants sorted and assigned to a fold by index modulo five, so no person is in training and test of the same fold", "labels": "the recording condition in the file name, pre-task rest against serial subtraction; the participants’ count-quality groups are left unused", "classes": [ "pre-task rest", "mental arithmetic" ] }, "model": { "id": "neurolm-b", "name": "NeuroLM-B", "family": "foundation", "paper_title": "NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals", "paper_venue": "ICLR 2025", "citation": "Jiang et al., ICLR 2025", "paper": "https://openreview.net/forum?id=Io9yFt7XH7", "paper_arxiv": "https://arxiv.org/abs/2409.00101", "repository": "https://github.com/935963004/NeuroLM", "source_revision": "0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177", "source_url": "https://github.com/935963004/NeuroLM/tree/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177", "checkpoint": { "file": "checkpoints/NeuroLM-B.pt", "repository": "https://huggingface.co/Weibang/NeuroLM", "revision": "eddfff5c64a4139442f826d6c67c8369fd00f45a", "sha256": "ffe098bc138b89f8817d3710a3604498d8ecd15135080e2ca27735d05c6d29ab", "bytes": 2377399148, "url": "https://huggingface.co/Weibang/NeuroLM/tree/eddfff5c64a4139442f826d6c67c8369fd00f45a" }, "parameters": 255107712, "restored_state_entries": 240, "parameters_basis": "a strict restore of the checkpoint, every state entry present and none unexpected", "architecture": { "layers": 12, "heads": 12, "width": 768, "context_tokens": 1024 }, "frozen": true, "fine_tuning": false, "text_generation": false, "language_model_head": false, "transfer": "supervised: the frozen NeuroLM outputs feed a linear head trained on the other folds’ labels", "weights_distributed": false, "weights_note": "no weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing", "endorsement": "none implied" }, "pretraining_exposure": { "status": "unknown", "status_vocabulary": [ "exposed", "not_exposed", "unknown" ], "checked_on": "2026-10-16", "published_pretraining_list": { "datasets": 16, "includes_eegmat": false, "source": "https://proceedings.iclr.cc/paper_files/paper/2025/file/8b4add8b0aa8749d80a34ca5d941c355-Paper-Conference.pdf" }, "instruction_tasks": [ "TUAB", "TUEV", "SEED", "HMC", "Workload", "TUSL" ], "eegmat_instruction_task": "Workload", "checkpoint_stage": { "status": "unknown", "checkpoint_epoch": 19, "iteration": 83660, "initialised_from": "gpt2", "learning_rate": 0.0006, "instruction_tuning_verified": false } }, "adapter": { "own_adapter": true, "source_sampling_hz": 500, "sampling_hz": 200, "resample": { "up": 2, "down": 5 }, "band_pass_hz": [ 0.1, 75 ], "filter_order": 4, "zero_phase": true, "offline": true, "filter": "Butterworth band-pass, applied forward and backward over the retained minute", "notch": false, "official_notch_hz": [ 50, 60 ], "demean": "per channel in each two-second window", "units": "microvolts, divided by 100 for NeuroLM; band power uses the microvolt windows", "amplitude_divisor": 100, "window_patches": 38, "patch_seconds": 1, "patch_order": "time-major, then channel" }, "probe": { "scaler": "StandardScaler fitted on training windows only", "head": "RidgeClassifier", "alpha": 100, "solver": "cholesky", "class_weight": "balanced", "fits": 15, "hyperparameter_search": false, "sealed_before_scoring": true }, "arms": [ { "id": "tokenizer_mean", "label": "tokenizer features", "dimension": 768, "definition": "the mean of NeuroLM-B’s 38 tokenizer hidden states per window (two one-second blocks of 19 channels), taken before its vocabulary projection", "balanced_accuracy": 0.6449074074074076, "macro_f1": 0.6251820691155802, "interval_95": null }, { "id": "log_bandpower", "label": "band power", "dimension": 95, "definition": "95 absolute log band powers: 19 channels by 5 bands (1–4, 4–8, 8–13, 13–30 and 30–75 Hz), Welch with a one-second Hann window at 200 Hz on the same windows", "balanced_accuracy": 0.6388888888888888, "macro_f1": 0.5907640266978917, "interval_95": null }, { "id": "gpt_mean", "label": "GPT hidden states", "dimension": 768, "definition": "the mean of the 38 final normalised GPT hidden states from the same EEG-only forward pass, which stops before the language-model head", "balanced_accuracy": 0.612037037037037, "macro_f1": 0.5933772131283214, "interval_95": null } ], "chance_level": { "balanced_accuracy": 0.5, "macro_f1": null }, "paired_contrasts": [ { "id": "tokenizer-minus-band-power", "a": "tokenizer_mean", "b": "log_bandpower", "label": "tokenizer features minus band power", "role": "question", "balanced_accuracy": { "difference": 0.00601851851851852, "interval_95": [ -0.05324074074074073, 0.0643865740740741 ], "interval_includes_zero": true, "reading": "the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent" }, "macro_f1": { "difference": 0.03441804241768836, "interval_95": [ -0.03926998065641698, 0.1067150716210463 ], "interval_includes_zero": true, "reading": "the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent" } }, { "id": "gpt-minus-tokenizer", "a": "gpt_mean", "b": "tokenizer_mean", "label": "GPT hidden states minus tokenizer features", "role": "descriptive", "balanced_accuracy": { "difference": -0.032870370370370376, "interval_95": [ -0.05371527777777777, -0.010173611111111137 ], "interval_includes_zero": false, "reading": "the interval lies below zero; descriptive, from this one cohort" }, "macro_f1": { "difference": -0.03180485598725867, "interval_95": [ -0.05423384240051305, -0.0074953822443153935 ], "interval_includes_zero": false, "reading": "the interval lies below zero; descriptive, from this one cohort" } }, { "id": "gpt-minus-band-power", "a": "gpt_mean", "b": "log_bandpower", "label": "GPT hidden states minus band power", "role": "descriptive", "balanced_accuracy": { "difference": -0.026851851851851863, "interval_95": [ -0.0833564814814815, 0.031030092592592557 ], "interval_includes_zero": true, "reading": "the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent" }, "macro_f1": { "difference": 0.0026131864304297, "interval_95": [ -0.06636081805766925, 0.0735941509659651 ], "interval_includes_zero": true, "reading": "the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent" } } ], "interval": { "kind": "paired participant bootstrap", "method": "percentile interval", "level": 0.95, "unit": "participant", "resamples": 1000, "seed": 20261016, "same_resamples_all_contrasts": true, "fixed_fits": true, "arms": "point estimates only" }, "separate_from_core": { "page": "/protocols/arithmetic-rest/", "shared_windows": true, "shared_channels": true, "shared_fold_rule": true, "shared_head": true, "core_band_pass": false, "core_sampling_hz": 250, "core_spectral_ridge": { "kind": "log relative spectrum", "band_hz": [ 0.5, 45 ], "sampling_hz": 250 }, "pilot_band_power": { "kind": "absolute log band power", "band_hz": [ 1, 75 ], "sampling_hz": 200 } }, "rights": { "dataset": { "name": "EEGMAT", "source": "https://physionet.org/content/eegmat/1.0.0/", "version": "PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0", "license": "Open Data Commons Attribution License 1.0", "licenseUrl": "https://opendatacommons.org/licenses/by/1-0/", "attribution": "Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z.", "privacyReview": "Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. EDF dates are normalized, but subject-info.csv contains age, gender, occupation, and recording date; those fields must not be published.", "reviewedAt": "2026-09-20", "reviewBasis": [ "https://physionet.org/content/eegmat/1.0.0/", "https://www.mdpi.com/2306-5729/4/1/14", "https://opendatacommons.org/licenses/by/1-0/" ], "rightsScope": "Personal noncommercial research; aggregate results only", "published_here": "Reused unchanged from the review of the same recordings for the core matrix. Published here for NeuroLM’s frozen features: equal-person balanced accuracy and macro-F1 of three representations, three paired differences with their participant-bootstrap intervals, and cohort counts; no per-person, per-fold or per-window value." }, "code": { "licence": "MIT", "scope": "NeuroLM’s source code at the pinned commit", "url": "https://github.com/935963004/NeuroLM/blob/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177/LICENSE", "licence_file_sha256": "21bddd47eec5ad9a56bc9bec45f91426680e373f957b840fde984252b6718185" }, "checkpoint": { "licence": "CC BY 4.0", "licence_url": "https://creativecommons.org/licenses/by/4.0/", "as_stated_by": "the licence tag in the Hugging Face repository metadata at the pinned revision", "model_card_bytes": 30, "model_card_terms": false, "gated": false, "checked_on": "2026-10-16", "weights_distributed": false }, "summary": "EEGMAT 1.0.0: Open Data Commons Attribution License 1.0 (Zyma et al., 2019; PhysioNet). NeuroLM code: MIT. NeuroLM-B checkpoint: CC BY 4.0 per its Hugging Face metadata; no weights distributed." }, "limitations": [ { "id": "exposure", "text": "EEGMAT is NeuroLM’s Workload task, one of the six instruction tasks in its pinned code, and the training stage of this checkpoint is unresolved: its metadata (epoch 19, iteration 83,660, initialised from GPT-2, learning rate 0.0006) fits a pretraining checkpoint, and instruction tuning is unverified. The backbone may have seen these people, so exposure is shown as unknown." }, { "id": "adapter", "text": "This pilot has its own offline adapter: a fourth-order zero-phase Butterworth band-pass at 0.1–75 Hz, resampling from 500 to 200 Hz, per-window demeaning, microvolts divided by 100 for NeuroLM, and no notch filter, where NeuroLM’s own pipeline adds a 50/60 Hz notch. The core Arithmetic & rest protocol filters nothing and resamples to 250 Hz, so these scores have their own panel." }, { "id": "band-power", "text": "The band-power arm is 95 absolute log band powers (19 channels by 5 bands, 1–75 Hz, Welch at 200 Hz), a different feature set from the core matrix’s Spectral ridge row, which uses log relative spectra from 0.5–45 Hz at 250 Hz." }, { "id": "recording-order", "text": "Rest is recorded before the task, so recording order moves with the label, and the pilot does not isolate cognitive workload." }, { "id": "capacity", "text": "The 95 band-power and 768 NeuroLM dimensions share one fixed alpha, which does not equalise capacity or effective regularisation." }, { "id": "random-initialisation", "text": "There is no random-initialisation control, so the pilot says nothing about a causal benefit of pretraining." }, { "id": "interval", "text": "Intervals come from 1,000 participant resamples of the 15 fixed fitted heads and exclude refit uncertainty. Each arm has a point estimate only." }, { "id": "development-cohort", "text": "EEGMAT is a known development cohort, used across this site. Predictions were sealed before scoring for procedural integrity: the labels are in the file names, and each person’s windows train the heads of the other folds." }, { "id": "ica", "text": "The source README reports prior ICA cleaning of the published signals." }, { "id": "scope", "text": "Frozen features with a supervised linear head: no text prompt, generated answer or language-model output entered any score, and nothing here measures instruction tuning, question answering, zero-shot use, cognitive workload, clinical use or online decoding." } ], "audits": { "independent_numerical_audit": { "status": "passed", "artifacts_hash_verified": 384, "refits": 15, "refits_exact": true, "saved_decisions_reconstructed": true, "bootstrap_max_abs_difference": 6.938893903907228e-18, "checkpoint_loads": 0, "sha256": "d329a085c87da98c138c1c03ce0042decafd0f72fca90514d5106a3e0da58d91" }, "reviews": { "execution": "985d6d7b74ec6c88455edeb7974a3c0f1ac18d65ffcefe3e0f1db9c045da32e6", "numerical_audit": "8454c6e776c7ab65882792c4ba60bbca7707eb0442c6842c92fddb8e42b94435", "interpretation": "cd28640d470dc298e27de270edb1768a1cf5128335e4ffc99ea40b2e99310e1c" }, "boundary": "The independent audit refitted all 15 heads from the sealed NeuroLM and band-power values and reproduced every score and interval; the NeuroLM features themselves were not re-extracted by a second implementation." }, "release_record": { "date": "2026-10-20", "decided": "2026-10-19", "decision": "Publish the NeuroLM EEGMAT frozen-feature pilot as its approved aggregates: a separate panel on the pilot studies page, outside the pilot count and every leaderboard.", "rights": "EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. NeuroLM’s code is MIT at the pinned commit. The NeuroLM-B checkpoint is CC BY 4.0 as the licence tag of its Hugging Face repository states, and its 30-byte model card adds no terms. No weights are distributed, and no endorsement by NeuroLM’s authors is implied.", "research_release": "The research session’s release record approves an aggregate pilot handoff, outside every leaderboard, and binds the export, protocol, producer, numerical audit and three reviews by SHA-256.", "research_release_status": "approved_for_aggregate_pilot_handoff", "binding": "This record approves the served file the export audit names." }, "not_published": [ "Per-person, per-fold and per-window values of any kind, fold assignments, participant identifiers, condition labels, features, fitted heads, predictions and the sealed prediction files: only equal-person means, paired differences with their intervals and cohort counts are published.", "Demographics from the source dataset’s participant file.", "Measured compute: worker times, memory and process counts from the handoff, the receipts and the audits.", "The checkpoint and the NeuroLM source: no weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing.", "The generation-mask adapter’s checks: engineering evidence that took no part in these scores.", "The IESSEEG intake and the watchlist leads of the same handoff: IESSEEG has its own release, and the leads have no results.", "The handoff, protocol, producer, receipts, reviews and input manifest themselves: each is pinned by SHA-256 in this manifest and checked by the export, and none is published, because several carry private storage paths." ], "provenance": { "manifest_sha256": "41c6c94772c50dcafb9624ad2d5ae20274a3f5b8d4617ef14a7e38105a45861e", "source_export_sha256": { "neurolm-eegmat": "62b7ebe4be2e331cb0651486185c7abdaa173d3279031d8ffc66c74b6208a20e" }, "research_release_sha256": "9a9c1b4770797c899b229e176f7deec4dceacd209c80dfcd3f63d2365afc7adf", "protocol_sha256": "670b55c1d63856af5b318104d40687772658f7790c9ebefd11f6633f4ed23b61", "numerical_audit_sha256": "d329a085c87da98c138c1c03ce0042decafd0f72fca90514d5106a3e0da58d91", "released_files_unchanged": [ "mvp.json", "pilot-studies-update.json", "eegpt-core-update.json" ] } }