All objects#
Every public object in the benchmark. Filter by name, module, or description.
Object |
Description |
|---|---|
|
The eight scored TS1 tasks: three sequence-level, five timestep-level. |
|
The two scored TS2 tasks, holding out units and trailing timesteps respectively. |
|
The scored TS3 tasks, named |
The flattened task id a submission is filed under, e.g. |
|
Whether a flattened task id belongs to |
|
Returns a list of all supported tasks. |
|
What a model needs to size a readout head, and nothing more. |
|
How many targets a task has along the input’s time axis: one, or one per timestep. |
|
What kind of value each target is, independent of |
|
Specification for a single decoding readout head. |
|
What |
|
What one TS2 task predicts. |
|
What |
|
What one TS3 task predicts. |
|
What |
|
|
Built-in immutable sequence. |
Raise unless |
|
Bits Per Spike metric for Poisson spiking data. |
|
Minimal implementation of Cohen pseudo-R^2 score for Poisson NLL loss. |
|
Compute and flatten a dict of torchmetrics metrics to plain Python scalars. |
|
Multi-unit averaging of prediction probabilities. |
|
Buffer tensors from an eval loop and save them as a |
|
The |
|
Pack string ids into an (n, width) uint8 tensor, width being the longest id. |
|
Unpack what |
|
Score one prediction file against its ground truth file. |
|
Score all TS1 prediction files found under |
|
Aggregate per-seed scores into mean ± SEM per (label, task, recording_id). |
|
Score one prediction file against its ground truth file. |
|
Score all TS2 prediction files found under |
|
Aggregate per-seed scores into mean ± SEM per (label, task, recording_id). |
|
Score one prediction file against its ground truth file. |
|
Score all prediction files found under |
|
Aggregate per-seed scores into mean ± SEM per label. |
|
Inject |
|
Clip designated metrics at |
|
Competition ranking, averaged over recording_id within each task. |
|
The filter provenance of a single recording. |
|
The filter provenance of a whole build, which may be mixed across recordings. |
|
Read the label of every recording the dataset covers. |
|
Raise unless the build satisfies the contract. There is no way to opt out. |
|
Read one vendored list, in file order. |
|
Dataset for the IBL BrainWideBench benchmark. |
|
Extends the sampled window backward for extra context. |
|
The benchmark sampled over whole sessions, after neural QC. |
|
Which sessions and units a whole-session dataset keeps. |
|
Base class for finetuning strategies. |
|
No-op strategy that leaves all model parameters trainable. |
|
Freeze all parameters except those matching unfrozen_prefixes, then unfreeze at a target epoch. |
|
Freeze all parameters except those matching unfrozen_prefixes. |
|
Standardized model interface. |
|
The epoch loop, and everything around it that depends on neither task nor model. |
|
Wrapper for distributing any sampler across multiple processes. |
|
Group any |
|
Drops units that have failing probe QC or did not fire in this slice. |
|
A simple unit dropout transform. |
|
Mask units and corresponding spikes given a boolean mask (in-place). |
|
Adds Gaussian noise scaled to the input max, clipped at 0. |
|
Randomly zeros out bins. |
|
Rescales amplitude by a small random factor. |
|
Adds Gaussian noise scaled to the input std. |
|
Apply a transform with probability p. |
|
Gaussian smoothing along the temporal (last) axis only. |
|
Random row-wise circular shift along the temporal axis. |
|
A simple extension of |
|
Linear readout with one head per task, routed by integer index. |
|
Scale a Transformer encoder’s weights in place for optimization without warmup. |
|
|
|
|
|
Narrow a config string to a known backend, raising on anything else. |
|
What to pass as |
|
Mean squared error under per-element weights, normalized by their sum. |
|
Symmetric InfoNCE loss over two paired views (CLIP, Eq. 1 of the NEMO paper). |
|
Effective rank of an |
|
Multi-session NDT with per-session stitchers [Ye and Pandarinath, 2021]. |
|
Temporal masking for spikes (BERT-style). |
|
Masked spike modelling for NDT Stitch, which fits one stitcher per session. |
|
Multi-task masked transformer over binned spikes [Zhang et al., 2024]. |
|
Module for masking spikes. Masking modes: |
|
Masked spike modelling for MtM, whose masks carry a mode token. |
|
Multi-context masked autoencoder over patched spike tokens [Ye et al., 2023]. |
|
ShuffleInfill-style MAE masker for NDT2. |
|
Masked spike modelling for NDT2, which masks context tokens and accumulates gradients. |
|
Multimodal masked transformer over spikes and behavior [Zhang et al., 2025]. |
|
Base class for all neural network modules. |
|
Multimodal masked modelling for NEDS, over spikes and the behaviors alike. |
|
Transformer-based model for neural decoding from spike trains [Azabou et al., 2023]. |
|
Supervised pretraining of POYO on one behavior. |
|
POYO+ over chained spike tokens [Azabou et al., 2025]. |
|
Supervised pretraining of POYO+ on every behavior at once, via its readout heads. |
|
POSSM (POYO + State Space Model) for the IBL benchmark [Ryoo et al., 2025]. |
|
Supervised pretraining of POSSM on one behavior. |
|
Multi-task pretrainer for POSSM. |
|
Reduced-rank decoder mapping a window of binned spikes to behavior [Zhang et al., 2026]. |
|
Supervised pretraining of the shared basis |
|
Contrastive unit encoder over population context [Arora et al., 2025]. |
|
Base class for all neural network modules. |
|
Contrastive pretraining of the NuCLR unit encoder, against |
|
NEMO bimodal contrastive learning model for IBL brain region pretraining [Yu et al., 2025]. |
|
Trainer for NEMO SSL pretraining (CLIP-style contrastive learning). |
|
Base class for all neural network modules. |
|
Base class for all neural network modules. |
|
Linear -> LayerNorm projection head (from NEMO LinearProjector, layer_norm=True). |
|
Binned spikes alone, for objectives that mask them and reconstruct them. |
|
Spikes aligned to one behavioral target. |
|
Spikes aligned to several behavioral targets at once. |
|
Dataset for the IBL BrainWideBench benchmark. |
|
Mixin enforcing the TS1 benchmark test contract. |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
Single linear layer mapping binned spike counts to task outputs. |
|
Multi-layer perceptron mapping binned spike counts to task outputs. |
|
Gated recurrent unit network mapping binned spike counts to task outputs. |
|
Temporal convolutional network mapping binned spike counts to task outputs. |
|
CEBRA-based neural decoding model on single-session [Schneider et al., 2023]. |
|
Benchmark eval trainer with in-run CEBRA encoder fitting. |
|
Single-session Transformer encoder for supervised neural decoding [Ye and Pandarinath, 2021]. |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
TS1 trainer for MtM, prompting the forward with an optional mask-mode token. |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
Standardized evaluation Trainer for Task Suite 1 (decoding). |
|
TS1 eval trainer plus the |
|
Dataset for the IBL BrainWideBench benchmark. |
|
Mixin enforcing the TS2 benchmark test contract. |
|
Standardized evaluation Trainer for Task Suite 2 (neural prediction). |
|
Simple Autoencoder with MLP encoder/decoder for TS2 neural prediction. |
|
LFADS for TS2 co-smoothing and forecasting [Keshtkaran et al., 2022, Pandarinath et al., 2018]. |
|
TS2 trainer for LFADS. |
|
Single-session Transformer encoder over binned spikes [Ye and Pandarinath, 2021]. |
|
TS2 trainer for the single-session NDT. |
|
Shared skeleton for every statistical baseline. |
|
Population coupling. |
|
Reduced-rank readout over binned counts. |
|
Reduced-rank readout over [counts, time-since-last-spike, last ISI]. |
|
Trailing-window mean, held flat over the horizon. |
|
Trial-local / global rate blend, held flat over the horizon. |
|
Per-unit ridge autoregression from the observed bins to the horizon. |
|
TS2 trainer for the zero-parameter statistical baselines. |
|
TS2 trainer for MtM, prompting the forward with the task’s mask-mode token. |
|
TS2 trainer for NDT Stitch. |
|
TS3’s view of the benchmark: whole sessions after neural QC. |
|
A unit-embedding producer, driven by |
|
ISI histogram, L1-normalized to sum=1. |
|
A unit-embedding producer, driven by |
|
A unit-embedding producer, driven by |
|
A unit-embedding producer, driven by |
|
A unit-embedding producer, driven by |
|
POYO+’s |
|
POSSM’s |
|
A unit-embedding producer, driven by |
|
A unit-embedding producer, driven by |
|
Masked spike modelling for NDT Stitch, which fits one stitcher per session. |
|
Masked spike modelling for MtM, whose masks carry a mode token. |
|
Multi-task pretrainer for POSSM. |
|
Supervised pretraining of POYO+ on every behavior at once, via its readout heads. |
|
ISI histogram encoder, attention pooling and a brain region classifier [Schneider et al., 2023]. |
|
Multi-Head Global pooling layer. |
|
Supervised training of LOLCAT, the one TS3 model fit on the region labels. |
|
Samples elements randomly from a given list of indices, without replacement. |
|
A classifier fit on unit embeddings, scored by |
|
Multinomial logistic regression on standardized embeddings. |
|
An MLP whose hyperparameters are swept with Optuna before the final fit. |