Available Models

This page provides an overview of the brain encoding models currently available in BERG.

Model Naming Convention

BERG contains several encoding models, defined by the following model ID naming convention:

{modality}-{dataset}-{model}

where

• modality: The neural recording modality on which the encoding model was trained.

• dataset: The neural dataset on which the encoding model was trained.

• model: The type of encoding model used.

For example:

• fmri-nsd-fwrf: An fMRI encoding model trained on the NSD using feature-weighted receptive fields.

• eeg-things_eeg_2-vit_b_32: An EEG model trained on the THINGS-EEG2 dataset using the ViT-B/32 visual transformer architecture.

Model Naming Convention (BrainScore models)

BERG additionally contains a special class of encoding models that are trained on BrainScore benchmarks. These models have a slightly different naming convention:

brainscore_{modality}-{model}

where

• modality: The neural recording modality of the BrainScore benchmark on which the encoding model was trained (i.e. vision, language).

• model: The specific BrainScore model used (e.g. AlexNet, GPT2-XL).

For example:

• brainscore_vision-alexnet: A vision model trained on the BrainScore vision benchmark using the AlexNet architecture.

• brainscore_language-gpt2: A language model trained on the BrainScore language benchmark using GPT-2.

Get Model Information

You can get detailed information about any model using:

from berg import BERG

berg = BERG("path/to/brain-encoding-response-generator")

# List all available models
all_models = berg.list_models()

# Get detailed model information
model_info = berg.describe("fmri-nsd-fwrf")

Available models

Following is a list of all available models, grouped by modality. The ✅ icon indicates the best model for each dataset.

modality-fmri

Encoding models trained on neural responses recorded with functional Magnetic Resonance Imaging (fMRI).

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	fmri-nsd_fsaverage-huze	Mapping of vision transformer image features onto fMRI responses.	Natural Scenes Dataset (surface space)	Human	Images	Accuracy plots
	fmri-nsd_fsaverage-vit_b_32	Linear mapping of vision transformer image features onto fMRI responses.	Natural Scenes Dataset (surface space)	Human	Images	Accuracy plots
	fmri-nsd_fsaverage-alexnet	Linear mapping of AlexNet image features onto fMRI responses.	Natural Scenes Dataset (surface space)	Human	Images	Accuracy plots
	fmri-nsd_fsaverage-alexnet_untrained	Linear mapping of an untrained AlexNet image features onto fMRI responses.	Natural Scenes Dataset (surface space)	Human	Images	Accuracy plots
✅	fmri-nsd-fwrf	Feature-weighted receptive fields, convolutional neural networks trained end-to-end to predict fMRI responses from input images.	Natural Scenes Dataset (volume space)	Human	Images	Accuracy plots
✅	fmri-bmd-s3d	Linear mapping of 3D CNN video features onto fMRI responses.	BOLD Moments Dataset (MNI152 volume space)	Human	Videos	Accuracy plots
✅	fmri-things_fmri_1-vit_b_32	Linear mapping of vision transformer image features onto whole-brain fMRI responses.	THINGS fMRI1	Human	Images	Accuracy plots
✅	fmri-mosaic-CNN8_multihead_subAll_verticesVisual	CNN predicting visual cortex responses (7,831 vertices) for 93 subjects across 8 datasets.	MOSAIC (all datasets)	Human	Images	Accuracy plots
✅	fmri-mosaic-CNN8_multihead_subNSD_verticesAll	CNN predicting whole-cortex responses (57,051 vertices) for 8 NSD subjects.	MOSAIC (NSD)	Human	Images	Accuracy plots
	fmri-tuckute_2024-GPT2_XL	GPT2-XL–based linear encoding model (LLM embeddings + ridge regression) predicting brain responses to sentences.	Tuckute et al., 2024	Human	Text	Accuracy plots
	fmri-cneuromod_algo2025-text2fmri	Transformer-based encoding model predicting whole-brain fMRI from video transcripts using LLM features	CNeuromod/Algonauts 2025	Human	Text	Hugging Face Collection
✅	fmri-cneuromod_algo2025-vibe	Multimodal fMRI encoding model fusing video, audio, and text features via a two-stage Transformer architecture	Algonauts 2025 Challenge	Human	Video + Audio + Text	Accuracy Plots
✅	brainscore_language	Gateway to 10+ GPT-family language models from BrainScore, mapped to human fMRI responses (9 subjects, 384 sentences) via PLS regression. Model weights hosted by BrainScore — not BERG.	Pereira et al., 2018	Human	Text	BrainScore leaderboard
✅	fmri-dascoli_2026-tribe_v2	Tri-modal (video, audio, language) Transformer encoding model predicting whole-brain fMRI (20,484 vertices). First place in Algonauts 2025 (263 teams).	Multi-study naturalistic fMRI	Human	Video + Audio + Text	TRIBE v2 paper
✅	fmri-lebel2023-opt_1_3b	OPT-1.3B–based linear encoding model (contextual LLM embeddings + ridge regression) predicting brain responses to spoken narratives, following Antonello et al., 2023.	LeBel et al., 2023	Human	Text	Accuracy plots

modality-eeg

Encoding models trained on neural responses recorded with Electroencephalography (EEG).

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	eeg-things_eeg_2-vit_b_32	Linear mapping of vision transformer image features onto EEG responses.	THINGS EEG2	Human	Images	Accuracy plots
	eeg-things_eeg_2-alexnet	Linear mapping of AlexNet image features onto EEG responses.	THINGS EEG2	Human	Images	Accuracy plots
	eeg-things_eeg_2-alexnet_untrained	Linear mapping of an untrained AlexNet image features onto EEG responses.	THINGS EEG2	Human	Images	Accuracy plots

modality-meg

Encoding models trained on neural responses recorded with Magnetoencephalography (MEG).

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	meg-things_meg_1-vit_b_32	Linear mapping of vision transformer image features onto time-resolved whole-brain MEG responses.	THINGS MEG1	Human	Images	Accuracy plots

modality-utah_array

Encoding models trained on neural responses recorded with Utah arrays (intracortical electrophysiology).

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	utah_array-tvsd-vit_b_32	Linear mapping of vision transformer image features onto time-resolved intracortical spiking activity.	THINGS Ventral Stream Spiking Dataset (TVSD)	Macaque	Images	Accuracy plots

modality-calcium_2p

Encoding models trained on neural responses recorded with two-photon calcium imaging.

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	calcium_2p-wang_2025-3DCNN	Foundation model of mouse visual cortex, based on a spatiotemporal convolutional neural network (3D CNN + ConvLSTM).	Wang et al., 2025	Mouse	Videos	Accuracy plots

modality-ephys

Encoding models trained on extracellular electrophysiology recordings from macaque visual cortex.

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	brainscore_vision	Gateway to 440+ vision models from BrainScore, mapped to macaque electrophysiology recordings (V1, V2, V4, IT) via PLS regression. Model weights hosted by BrainScore — not BERG.	Freeman et al., 2013 (V1, V2); Majaj et al., 2015 (V4, IT)	Macaque	Images	BrainScore leaderboard

modality-ecog

Encoding models trained on neural responses recorded with electrocorticography (ECoG).

Best model	Model ID	Description	Training dataset	Species	Stimuli	Encoding accuracy
✅	ecog-zada2025-gpt2_xl	Linear mapping of GPT-2 XL contextual word embeddings onto time-resolved high-gamma ECoG activity during natural speech comprehension.	Zada et al., 2025	Human	Text	Accuracy plots