IBM-2 research checkpoints (public subset, 2026-10-08; revised 2026-10-10)
Withdrawn 2026-10-10. The Yeo-2011-dependent fMRI checkpoints (351 files: the fMRI masked-infill diagnosis P1–P8, field comparison v1/v2 fMRI, field scaling v1 fMRI and the fMRI infill pilot) were removed from the current revision pending a redistribution licence for the Yeo-2011 parcellation, whose six network centroids they embed in their config. They are kept privately in the IBM-2 archive. The 23 remaining checkpoints are trained only on synthetic data generated in the codebase.
IBM-2 is a research codebase that builds an anatomically structured brain/body model and evaluates it on neural-prediction, stimulation-response and body tasks under preregistered protocols. This repository holds the subset of IBM-2 training checkpoints whose training data permits redistribution of derived models.
These are research checkpoints, not validated models. No biological validity, clinical utility or benchmark standing is claimed. The remaining checkpoints are synthetic engineering smokes, demos and synthetic gates, which ended negative or were not scored. Results below are quoted from the project's committed workstream records and apply only to the stated protocol.
What is included and why
Only checkpoints whose every training input has a recorded licence permitting redistribution were released; after
the 2026-10-10 withdrawal that is purely synthetic data generated in the codebase. Checkpoints trained on data
with unresolved, non-commercial, share-alike-unclear, DUA or "local-only" terms were not uploaded (they include
EEGMMIDB, Sleep-EDF, THINGS, LibriBrain, OSF wsgzp, Midnight Scan Club, NLB/DANDI, body-model assets, and every model
that embeds the BrainGraph/DK83 atlas). No datasets, held-out or sealed data, or predictions are included: only
checkpoint files, this card and manifest.json (sha256 per file).
None of these checkpoints was trained on held-out or sealed test data.
Format. Each .pt is a torch.save dict with state_dict, optimizer, config, config_sha256, provenance,
progress and RNG state. It needs the IBM-2 source at the listed commit to rebuild the module. Loading requires
torch.load(path, weights_only=False) (pickle): only do this if you trust this repository. Embedded configs may contain
original absolute local paths. See manifest.json for per-file sha256.
Folders
| Folder | Files | Protocol | Source commit | Training data (licence) | Recorded result |
|---|---|---|---|---|---|
checkpoints (legacy synthetic demos) |
3 | synthetic only (in-repo demo anatomy, elbow-reach fixture) | repository import ab5ec8c | synthetic (n/a (generated in-repo)) | not scored (smoke/demo) |
out/diagnostics/unified-environment-01 |
1 | synthetic only (toy anatomy, elbow reach, synthetic TMS targeting; no human data) | not recorded | synthetic (n/a (generated in-repo)) | not scored (2-step engineering smoke) |
out/diagnostics/unified-session-01 |
2 | synthetic only (toy anatomy, elbow reach, synthetic TMS targeting; no human data) | not recorded | synthetic (n/a (generated in-repo)) | not scored (2-step engineering smoke) |
out/field-comparison-v2/synthetic |
15 | Field comparison v2 synthetic gate (short bias / ablated / random-null-key factor, 3 seeds) | fdfb163 (v2 prereg) | synthetic (in-repo covariance_contrast_fixture, src/ibm/experiments/supervised_diagnosis.py) (generated in-repo; no third-party input) | negative: no zero-null FT arm passes; short bias > ablated (0.658 vs 0.508); random null key 0.867 |
out/field-comparison/synthetic-ft_ablated |
1 | Field comparison stage 3 (prereg v1/v1.1) synthetic covariance gate, FT ft_bias/ft_ablated | 6098709 (prereg v1), b05d2cf (v1.1) | synthetic (in-repo covariance_contrast_fixture, src/ibm/experiments/supervised_diagnosis.py) (generated in-repo; no third-party input) | ft_bias held-out 0.90 (borderline, float flag fail), ft_ablated 0.50 (memorizes); single seed |
out/field-comparison/synthetic-ft_bias |
1 | Field comparison stage 3 (prereg v1/v1.1) synthetic covariance gate, FT ft_bias/ft_ablated | 6098709 (prereg v1), b05d2cf (v1.1) | synthetic (in-repo covariance_contrast_fixture, src/ibm/experiments/supervised_diagnosis.py) (generated in-repo; no third-party input) | ft_bias held-out 0.90 (borderline, float flag fail), ft_ablated 0.50 (memorizes); single seed |
Licence of these checkpoint files: CC0-1.0.