Published Foundation Models, One Interface
neuros-foundation wraps the field's reference neural foundation models behind a shared
base_foundation_model contract, so they can be loaded, run, and compared like any other neurOS
component.
package map
neuros-foundation — reference neural foundation models
Published neural foundation models behind one base interface, with dataset loaders, so external architectures are directly comparable inside neurOS.
Models
CEBRA
consistent latent embeddings
NDT
neural data transformer
POYO / Neuroformer
spiking and multimodal foundation models
Interface
base_foundation_model
shared load / infer contract
utils
adapters and helpers
Datasets
Allen datasets
atlas-scale loaders
BCI datasets
benchmark tasks
What It Comprises
- CEBRA — consistent latent embeddings of neural and behavioral data.
- NDT — the Neural Data Transformer.
- POYO — a spiking / population foundation model.
- Neuroformer — a multimodal neural foundation model.
- Interface —
base_foundation_modelplusutilsfor a shared load/infer path. - Datasets — Allen and BCI dataset loaders for benchmarking.
How It Differs
This package is about reusing and benchmarking published architectures, not building one. It is
the comparison harness that sits next to neuros-neurofm (my own neuroFMx model) — so I can measure
a from-scratch design against established baselines on the same loaders and interface.
Part of the neurOS-v1 platform.