IBM-1 — implicit brain model checkpoints

Weights for github.com/JacobFV/IBM-1: a brain declared once, as four primitives, from which task-specific models are materialised on demand. An EEG forward model, a hemodynamic model and an audio-visual predictor are different projections of one parameter set.

These are research checkpoints from an in-progress programme. They do not yet demonstrate task performance — stimulus→brain skill currently hovers at chance, and nothing here is replicated across seeds. They are published so results are reproducible against the declaration that produced them, not as a usable model.

Checkpoint naming

A step number identifies nothing here: the substrate changes, the ontology changes, and 32M association embeddings are meaningless without knowing how many sites they index and what dynamics ran between them. So the name carries all of it.

ibm1.m-{modality}.s{sites}.e{embed}.k{degree}.dyn-{rev}.obj-{objective}.vw{viability}.step{n}.git-{sha}

field	meaning
m	v video · a audio · av joint · p paired-neural · multi
s / e / k	site count · embedding dim · association degree
dyn	dynamics revision — the field that catches "the weights load and the model is wrong"
obj	what was minimised
vw	−log₁₀ viability weight
git	the ontology commit the weights were trained against

k matters: it sets the fan-in normalisation, so two checkpoints of identical shape and different k carry differently scaled weights. Each .pt ships a .json sidecar with geometry, parameter split and metrics at that step.

What is trained

Only the learned factor of the declared association weight:

w_ij = M[parcel] × exp(−d_ij/ℓ) × σ(⟨e_i, e_j⟩)
       tractography  geometry     LEARNED, per site

Tractography constrains the coarse modular graph; the fine cortico-cortical graph is learned. Geometry and the parcel-scale prior are held.

Verified property

Ablating a trained checkpoint: bypassing the cortical dynamics costs +324% loss, freezing the learned association weights +180%. The substrate is load-bearing, not a delay line.

Data and attribution

Trained against openly licensed and research-licensed corpora catalogued in the repository, each card naming its origin. Datasets with attribution requirements are cited there, including THINGS (Hebart et al., PLoS ONE 2019; Stoinski et al., Behav Res 2024), used for research under a non-commercial licence, neither redistributed nor altered.

Licence cc-by-nc-4.0 reflects the most restrictive term inherited from the training corpora.