fMRI基础模型的扩展性研究

# A Scaling Study for fMRI Foundation Models
Source: [https://arxiv.org/html/2609.27232](https://arxiv.org/html/2609.27232)
Xuanye PanJunfeng XiaJunxiang ZhangAffiliation:University of Warwick SLAIMo WangAffiliation:Equal authorQuanying LiuAffiliation:Equal author

###### Abstract

Scaling laws have guided large\-model development in computer vision and natural language processing, but the relationships among data, model size, and compute remain unclear for functional magnetic resonance imaging \(fMRI\) foundation models\. Here, we conduct a controlled empirical study using pretraining data from more than 200 source datasets and over 10,000 GPU\-hours of experiments\. Holding the pretraining framework and downstream protocol fixed, we vary pretraining data size, model size, and training duration\. Downstream performance generally improves with compute, yet models using similar compute can perform substantially differently\. Additional pretraining data bring larger gains at larger model sizes, suggesting that data and model size should be scaled together\. At matched compute, increasing pretraining data benefits more tasks than increasing model size, although the pattern varies across tasks\. We then use in\-distribution \(ID\) downstream performance to select the combination of pretraining data size, model size, and training duration at two fixed compute budgets\. The resulting models are locked before out\-of\-distribution \(OOD\) evaluation\. They achieve the highest average performance across the evaluated OOD tasks among the compared fMRI foundation models while using less pretraining compute\. Overall, our results show that compute alone does not characterize fMRI scaling: performance depends on how pretraining data, model size, and training duration are combined\. Code is available at[link](https://github.com/derrz2/neurojepa)\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/scaling_fig1.jpg)Figure 1:Compute allocation shapes downstream performance\.Each trajectory follows one data\-model configuration; marker area denotes pretraining data size and color denotes model size\. At comparable compute, configurations reach different mean linear\-probe errors across 12 tasks; within the evaluated range, data gains are larger for larger models\.## 1Introduction

Scaling studies in language and vision have provided practical guidance for relating data, model size, and compute to performance\([Kaplan et al\., 2020](https://arxiv.org/html/2609.27232#bib.bib8);[Hoffmann et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib6);[Zhai et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib29);[Cherti et al\., 2023](https://arxiv.org/html/2609.27232#bib.bib2)\)\. Functional magnetic resonance imaging \(fMRI\) operates under a different constraint: pretraining relies on a finite collection of human recordings, while labeled downstream datasets are often small\([Marek et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib11);[Ortega Caro et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib186);[Gijsen et al\., 2026](https://arxiv.org/html/2609.27232#biba.bib14);[Wang et al\., 2025](https://arxiv.org/html/2609.27232#bib.bib20);[Xia et al\., 2026b](https://arxiv.org/html/2609.27232#bib.bib26)\)\. More pretraining data, a larger model, and continued training can consume similar compute without necessarily producing the same downstream benefit\. The central question is therefore how pretraining data size, model size, and training progress jointly shape downstream performance\.

Existing fMRI foundation\-model studies provide encouraging evidence that larger pretraining datasets or larger models can improve downstream performance\([Ortega Caro et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib186);[Gijsen et al\., 2026](https://arxiv.org/html/2609.27232#biba.bib14);[Wei et al\., 2026](https://arxiv.org/html/2609.27232#biba.bib199);[Yang et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib202);[Dong et al\., 2025](https://arxiv.org/html/2609.27232#biba.bib11);[Wang et al\., 2026b](https://arxiv.org/html/2609.27232#bib.bib22)\)\. However, these factors have largely been examined separately or only at the end of training\. Consequently, it remains unclear whether additional data become more useful as models grow, whether this relationship changes during training, and whether different data–model configurations remain comparable when they use similar compute\. We address these questions through a controlled study centered on downstream performance\.

To isolate scaling behavior, we hold the pretraining objective, architecture family, and downstream protocol fixed\. All models use NeuroJEPA, a simple ROI\-time implementation following LeJEPA\([Balestriero and LeCun, 2025](https://arxiv.org/html/2609.27232#bib.bib1)\), as a common pretraining method rather than as a scaling variable\. Our experiments use data assembled from more than 200 source datasets\. We follow each supported data–model configuration through multiple stages of training and evaluate frozen representations on 12 downstream tasks\. This design lets us compare continued training within one configuration with differences across configurations using similar estimated compute\.

Three patterns emerge\. First, downstream performance generally improves with additional training, but models using similar compute can still perform differently\. Second, additional pretraining data are associated with larger gains for larger models, and this relationship becomes clearer later in training\. Third, at matched compute, increasing pretraining data provides the more consistent gain across downstream tasks, whereas the benefit of increasing model size is more task dependent\. These results show that fMRI scaling cannot be characterized by compute alone: pretraining data size, model size, and training compute need to be considered jointly\.

Finally, we fit in\-distribution \(ID\) downstream performance over the observed combinations of pretraining data size, model size, and training duration\. At each of two prespecified compute budgets, the fitted relationship selects the combination with the best predicted ID performance\. We lock both selections before examining any out\-of\-distribution \(OOD\) results and then compare the selected models with released fMRI foundation models\.

We make three contributions:

- •We conduct a controlled, scaling study of fMRI foundation models\. Using data assembled from more than 200 source datasets and over 10,000 GPU\-hours of experiments, we evaluate six nested pretraining data sizes, model sizes ranging from approximately 2M to 92M encoder parameters, and multiple compute budgets on 12 frozen linear\-probe tasks under one fixed protocol\.
- •We characterize how pretraining data and model size scale together throughout training\. By following each supported data–model configuration over time, we show that larger models are associated with greater gains from additional pretraining data\.
- •We compare data scaling and model scaling task by task at matched compute\. Most tasks show a clearer benefit from additional pretraining data, while the benefit of increasing model size is more task dependent\. We then use ID downstream performance to select two models under fixed compute budgets and evaluate the locked models on OOD tasks\.

## 2Related Work

##### Scaling data, model size, and compute\.

Scaling studies in language and vision vary data and model size over several orders of magnitude, fit predictable loss or performance trends, and use them to compare compute\-efficient training configurations\([Kaplan et al\., 2020](https://arxiv.org/html/2609.27232#bib.bib8);[Hoffmann et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib6);[Zhai et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib29);[Cherti et al\., 2023](https://arxiv.org/html/2609.27232#bib.bib2)\)\. Although data constraints are now also recognized in language\-model training\([Muennighoff et al\., 2023](https://arxiv.org/html/2609.27232#bib.bib12)\), fMRI begins in a more acquisition\-constrained regime\. Additional pretraining data require human recordings, while reproducible brain–phenotype associations can require thousands of participants even though typical neuroimaging cohorts are much smaller\([Marek et al\., 2022](https://arxiv.org/html/2609.27232#bib.bib11)\)\. Expanding an fMRI corpus also commonly adds participants, sites, scanners, and acquisition protocols, so its data axis changes both the amount and composition of the available evidence\([Wang et al\., 2026a](https://arxiv.org/html/2609.27232#bib.bib21);[Xia et al\., 2026a](https://arxiv.org/html/2609.27232#bib.bib25)\)\. Moreover, the relevant outcome is downstream performance across label\-limited tasks, not pretraining loss alone\([Xia et al\., 2026c](https://arxiv.org/html/2609.27232#bib.bib27)\)\. We therefore study how pretraining data, model size, and estimated compute jointly relate to downstream performance, and how these relationships change with training progress and across tasks\.

##### Scaling fMRI pretraining\.

BrainLM introduced explicit data\- and model\-size experiments into fMRI foundation modeling, reporting masked\-signal reconstruction and limited downstream comparisons across pretraining scales\([Ortega Caro et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib186)\)\. Brain\-Semantoks studied data scaling with a fixed architecture across frozen linear probes\([Gijsen et al\., 2026](https://arxiv.org/html/2609.27232#biba.bib14)\)\. These studies show favorable trends along individual scaling axes, but do not jointly characterize how pretraining data, model size, and training progress should be allocated under a common compute budget\. CortexMAE provides the closest systematic study: it performs data\- and model\-size sweeps on HCP\-YA, fits power\-law trends for masked\-reconstruction loss, and reports downstream trends on four selected targets\([Lane et al\., 2025](https://arxiv.org/html/2609.27232#bib.bib9)\)\. It also examines performance across training progress, but does not compare alternative data–model allocations at matched total compute\. We instead relate downstream performance to nested multi\-study pretraining data sizes, model size, and training progress\.

##### Self\-supervised objectives for fMRI\.

Early fMRI foundation models primarily learned by reconstructing masked BOLD signals\. BrainLM reconstructs ROI\-time patches, while CortexMAE applies masked reconstruction to cortical flat maps\([Ortega Caro et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib186);[Lane et al\., 2025](https://arxiv.org/html/2609.27232#bib.bib9)\)\. BrainMass is a hybrid masked\-modeling and latent\-alignment method for functional\-connectivity matrices\. It constructs two pseudo\-functional connectivity views by dropping BOLD time points, encodes them with online and EMA\-updated target networks, and aligns their normalized latent embeddings with an online predictor\. This alignment is trained jointly with masked\-ROI identification and reconstruction losses\([Yang et al\., 2024](https://arxiv.org/html/2609.27232#biba.bib202)\)\. Brain\-Semantoks learns temporally stable representations through self\-distillation between temporal views\. It combines an EMA teacher–student pair with a functional\-network tokenizer, masked\-token prediction, coding\-rate regularization, and an early\-training curriculum\([Gijsen et al\., 2026](https://arxiv.org/html/2609.27232#biba.bib14)\)\. For controlled scaling, we use a deliberately simpler pretraining method\. Following LeJEPA\([Balestriero and LeCun, 2025](https://arxiv.org/html/2609.27232#bib.bib1)\), NeuroJEPA aligns global and local views of the same ROI\-time segment with a shared encoder and applies SIGReg to encourage isotropic latent geometry\. Compared with BrainMass and Brain\-Semantoks, it uses no EMA target network; it also requires no masked reconstruction head, semantic tokenizer, or training curriculum\. We do not claim that this objective is universally optimal; it provides a fixed pretraining method for studying how pretraining data, model size, and training progress relate to downstream performance\.

## 3Fixed Pretraining Method

This section defines the representation learner used in every scaling experiment\. NeuroJEPA applies the LeJEPA objective\([Balestriero and LeCun, 2025](https://arxiv.org/html/2609.27232#bib.bib1)\)to ROI time series\. The purpose of this section is to make the fixed pretraining method clear before the scaling variables are introduced\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/lejepa.jpg)Figure 2:NeuroJEPA pretraining, illustrated with one of two global views and one of six local views\. All views share the same encoder and projection head\. The loss aligns every view to the mean global representation and applies isotropic geometry regularization to each view batch\.An input segment is a matrix of ROI signals over time\. For a minibatch ofBBsegments, letxi∈ℝR×Tx\_\{i\}\\in\\mathbb\{R\}^\{R\\times T\}denote segmentii, whereRRis the number of cortical ROIs \(R=100R=100in all experiments\) andTTis the number of time points\. We constructV=8V=8augmented views of each segment: two*global views*that retain broad ROI\-time context and six*local views*that retain less context\. Thevvth view of segmentiiisxi\(v\)x\_\{i\}^\{\(v\)\}\. All views come from the same input segment; the complete sampling recipe is in the supplementary material\.

The shared encoderfθf\_\{\\theta\}is a Transformer over small blocks of the ROI\-by\-time matrix\. It maps each view to one pooled feature vector\. A shared projection headgϕg\_\{\\phi\}, a small network applied after the encoder, maps that feature vector to thedzd\_\{z\}\-dimensional embeddingzi\(v\)=gϕ​\(fθ​\(xi\(v\)\)\)∈ℝdzz\_\{i\}^\{\(v\)\}=g\_\{\\phi\}\(f\_\{\\theta\}\(x\_\{i\}^\{\(v\)\}\)\)\\in\\mathbb\{R\}^\{d\_\{z\}\}\. Let𝒢\\mathcal\{G\}be the set of global\-view indices\. The mean global embedding for segmentiiis

ci=1\|𝒢\|​∑v∈𝒢zi\(v\)\.c\_\{i\}=\\frac\{1\}\{\|\\mathcal\{G\}\|\}\\sum\_\{v\\in\\mathcal\{G\}\}z\_\{i\}^\{\(v\)\}\.All global and local embeddings are aligned to this mean:

ℒalign=1B​V​∑i=1B∑v=1V‖zi\(v\)−ci‖22\.\\mathcal\{L\}\_\{\\mathrm\{align\}\}=\\frac\{1\}\{BV\}\\sum\_\{i=1\}^\{B\}\\sum\_\{v=1\}^\{V\}\\left\\\|z\_\{i\}^\{\(v\)\}\-c\_\{i\}\\right\\\|\_\{2\}^\{2\}\.Gradients pass through the mean and every view embedding; no separate target encoder is used\.

The second term is Sketched Isotropic Gaussian Regularization \(SIGReg\) from LeJEPA\. For one view indexvv, SIGReg projects the batch\{zi\(v\)\}i=1B\\\{z\_\{i\}^\{\(v\)\}\\\}\_\{i=1\}^\{B\}onto random one\-dimensional directions and compares each projected distribution with a standard Gaussian using the Epps–Pulley statistic\. This discourages collapsed or strongly concentrated embeddings\. Letλ∈\(0,1\)\\lambda\\in\(0,1\)be its loss weight\. The complete objective is

ℒNeuroJEPA=\(1−λ\)​ℒalign\+λV​∑v=1VℒSIGReg​\(\{zi\(v\)\}i=1B\)\.\\mathcal\{L\}\_\{\\mathrm\{NeuroJEPA\}\}=\(1\-\\lambda\)\\mathcal\{L\}\_\{\\mathrm\{align\}\}\+\\frac\{\\lambda\}\{V\}\\sum\_\{v=1\}^\{V\}\\mathcal\{L\}\_\{\\mathrm\{SIGReg\}\}\\left\(\\\{z\_\{i\}^\{\(v\)\}\\\}\_\{i=1\}^\{B\}\\right\)\.We use the LeJEPA SIGReg functional unchanged; the supplementary material states the implementation and hyperparameters\. The intuition is simple: alignment preserves information shared across different views of the same segment, while SIGReg keeps the representation well spread for later linear probes\. The longer geometry argument is placed in the appendix because the paper’s main question is scaling\. We do not claim that this objective preserves every downstream variable or is optimal for every fMRI task\.

With the pretraining method fixed, Section[4](https://arxiv.org/html/2609.27232#S4)defines the pretraining data size, model size, training progress, and estimated compute used below\.

## 4Controlled Scaling Setup

We hold the NeuroJEPA objective, ROI\-time input, view construction, architecture family, optimizer family, and downstream protocol fixed\. We vary pretraining data size, model size, and progress through the planned training schedule\.

### 4\.1Data, Models, and Compute

Dataset / sourceDownstream taskSubjectsMetricsIn\-distribution \(ID\)ABIDE\([Di Martino et al\., 2014](https://arxiv.org/html/2609.27232#biba.bib9)\)Autism diagnosis \(binary\)871Accuracy / macro\-F1ABIDEAge prediction871MSE / PearsonrrPNC\([Satterthwaite et al\., 2014](https://arxiv.org/html/2609.27232#biba.bib191)\)Sex classification \(binary\)1,268Accuracy / macro\-F1PPMI\([Marek et al\., 2011](https://arxiv.org/html/2609.27232#biba.bib23)\)Diagnosis classification \(3\-way\)474Accuracy / macro\-F1HCP\([Van Essen et al\., 2013](https://arxiv.org/html/2609.27232#biba.bib197)\)Sex classification \(binary\)1,010Accuracy / macro\-F1Out\-of\-distribution \(OOD\)ADNI\([Jack et al\., 2008](https://arxiv.org/html/2609.27232#biba.bib19)\)Alzheimer’s disease vs\. control230Accuracy / macro\-F1ADNI†\\daggerMild cognitive impairment vs\. control292Accuracy / macro\-F1ADHD\-200\([The ADHD\-200 Consortium, 2012](https://arxiv.org/html/2609.27232#biba.bib196)\)ADHD diagnosis \(binary\)696Accuracy / macro\-F1BHRC\([Salum et al\., 2025](https://arxiv.org/html/2609.27232#biba.bib190)\)Sex classification \(binary\)465Accuracy / macro\-F1NKI\-RS\([Nooner et al\., 2012](https://arxiv.org/html/2609.27232#biba.bib27)\)Age prediction717MSE / PearsonrrNKI\-RSEducation classification \(3\-way\)717Accuracy / macro\-F1SALD\([Wei et al\., 2018](https://arxiv.org/html/2609.27232#biba.bib198)\)Age prediction492MSE / PearsonrrTable 1:Downstream datasets and evaluation tasks\.Counts denote unique participants after task\-specific filtering and deduplication across the training, validation, and test splits\. Detailed task descriptions and metric construction are provided in the supplementary material\.The pretraining data contain fMRI signals from more than 200 source datasets\. To obtain the common ROI\-time series, we spatially resample each fMRI recording to 2 mm isotropic resolution and temporally resample it onto a common model\-input grid with 0\.72 s spacing\. We then use the 100\-parcel Schaefer cortical parcellation\([Schaefer et al\., 2018](https://arxiv.org/html/2609.27232#bib.bib17)\)to extract a 100\-channel cortical ROI\-time series\. All resulting ROI\-time series undergo the same temporal normalization\. The main crossed grid contains pretraining data sizes of 2,585, 3,810, 6,533, 11,673, 32,414, and 63,401 recordings\. All levels use the same source datasets and differ only in the participant fraction sampled within each source, preserving source proportions up to rounding\. They are subject\-grouped and strictly nested: participants enter with all recordings, and every smaller level is contained in the next larger level\.

LetDDdenote*pretraining data size*, measured by the number of distinct recordings\. Reusing a recording later in training does not increaseDD\.

LetNNdenote*model size*, measured by the number of trainable encoder parameters and excluding the projection head\. The encoder is a Vision Transformer over ROI\-time patches, and the model family varies its width and depth\. The main scaling analyses cover model sizes from approximately 2 million to 92 million encoder parameters\. Each analysis reports its trained support rather than implying a complete Cartesian grid\. The supplementary material lists the architectures, and the supplementary dataset table reports the recordings and participants from each source\.

LetCCdenote cumulative training compute in EFLOP, where1​EFLOP=10181\\ \\mathrm\{EFLOP\}=10^\{18\}floating\-point operations\. We estimateCCfrom the dominant dense\-matrix operations in the encoder, the projection head, and the number of optimizer steps; the supplementary material gives the accounting\. For training runii, letssbe the number of completed optimizer steps andSiS\_\{i\}its planned total\. Training progress isp=s/Sip=s/S\_\{i\}, sop=0\.5p=0\.5is the midpoint of that run’s schedule andp=1p=1is its endpoint\. Equalppdoes not imply equal compute across runs\. For the allocation analysis in Section[6](https://arxiv.org/html/2609.27232#S6), leteedenote*epoch\-equivalent exposure*: the cumulative number of recording presentations divided byDD\. Thus,e=1e=1corresponds to one average pass over the available pretraining recordings\. Unlikepp, which measures progress relative to a run’s planned schedule,eemeasures how often the available recordings have been presented\.

### 4\.2Evaluation Protocol

The 12 fixed tasks span demographic, clinical, with both classification and regression\. They cover sex, age, education, autism, Parkinson’s disease, Alzheimer’s disease, mild cognitive impairment, and ADHD across ABIDE, PNC, PPMI, SALD, ADNI, ADHD\-200, BHRC, HCP, and NKI\-RS\. Table[1](https://arxiv.org/html/2609.27232#S4.T1)lists the task groups, participant counts, and metrics\. In Figure[3](https://arxiv.org/html/2609.27232#S5.F3), classification and regression group tasks by target type, while in\-distribution \(ID\) and out\-of\-distribution \(OOD\) denote two fixed dataset groups listed in that table\.

Subjects follow 6:2:2 train, validation, and test partitions\. All recordings from one participant remain in one split\. The train partitions remain fixed across five evaluation seeds\. The pretraining data exclude downstream validation and test participants\. For each task, the encoder is frozen, and a linear probe is trained for each downstream task\. For cross\-task plots, a classification score averages macro\-F1\. A regression score averages Pearson correlation\. Each score is oriented so that higher is better, averaged across the five probe seeds\. The plotted error is one minus score; the supplementary material gives the exact construction\.

## 5How Should fMRI Pretraining Scale?

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/scaling_fig3.jpg)Figure 3:Training trajectories and data–model\-size relationships during training\.\(a\) Linear\-probe error for fixed data–model configurations\. \(b\) Standardized downstream scores at matched fractions of each training progress; circles are evaluated configurations and contours are descriptive interpolations\. \(c\) Task\-resampling uncertainty forβD×N\\beta\_\{D\\times N\}; positive values mean that larger models show larger fitted gains from additional pretraining data\.We organize the scaling analysis around training progress, the joint data–model relationship, and comparisons at similar estimated compute\. The analyses use higher\-is\-better, outer\-validation scores\. The data axis is the nested pretraining data sequence defined in Section[4](https://arxiv.org/html/2609.27232#S4)\. We describe this observed grid rather than fit or extrapolate a universal scaling law\. Panel \(a\) shows the equivalent lower\-is\-better error; the remaining analyses use the higher\-is\-better score\.

### 5\.1Compute Helps, but Scaling Choices Matter

Figure[3](https://arxiv.org/html/2609.27232#S5.F3)\(a\) follows the task\-group\-averaged linear\-probe error of each fixed data–model configuration as training progresses\. Across classification, regression, ID, and OOD, error generally decreases as estimated compute accumulates\. The pattern is not strictly monotonic, but its overall direction is consistent across all four task\-group summaries\.

At the upper end of the observed compute range, configurations combining larger pretraining data sizes with larger model sizes increasingly occupy the lower\-error region\. This is a descriptive pattern within the evaluated grid: pretraining data size, model size, training progress, and estimated compute change together, so Figure[3](https://arxiv.org/html/2609.27232#S5.F3)\(a\) does not isolate the independent effect of any one factor\.

Estimated compute nevertheless does not uniquely order the observed configurations\. Trajectories with overlapping or nearby compute ranges can remain at different error levels and sometimes change their ordering during training\. Compute therefore tracks the amount of training but does not fully describe how efficiently that compute is used\. This motivates separating the data\-size and model\-size axes and asking whether their relationships with downstream performance depend on one another\.

### 5\.2Data Gains Depend on Model Size

We next ask a concrete question: is additional pretraining data associated with the same performance change for small and large models? Figure[3](https://arxiv.org/html/2609.27232#S5.F3)\(b\) maps downstream performance over the observed data–model grid at matched stages of training progress\. If the two resources were unrelated, moving toward more data would produce a similar fitted change on every model\-size row\. Instead, the fitted relationship with data size becomes more favorable in larger\-model rows\. Equivalently, the relationship with model size becomes more favorable when more pretraining data are available\.

To check whether this visual pattern is shared across tasks, we measure, for each task, how the fitted gain from additional pretraining data changes as model size increases\. We average these changes over the 12 tasks and denote the result byβD×N\\beta\_\{D\\times N\}\. A positive value has a direct interpretation: larger models show larger fitted gains from additional pretraining data\. A value near zero would mean that the fitted data\-size relationship is similar across model sizes\.

Figure[3](https://arxiv.org/html/2609.27232#S5.F3)\(c\) repeatedly resamples the task set to show the uncertainty inβD×N\\beta\_\{D\\times N\}\. It is positive throughout the measured training stages and larger at the end of training than at the midpoint\. In practical terms, increasing model size and increasing pretraining data size are most effective when they are scaled together and trained sufficiently\. This result applies to the evaluated grid and does not establish a causal or strictly monotonic relationship\. The supplementary material provides the regression definition, numerical estimates, and sensitivity analyses\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/fig4.jpg)Figure 4:Task\-wise comparison of scaling pretraining data and model size at matched compute\.\(a\) Each task is positioned by the performance changes associated with increasing data size and model size; the diagonal denotes equal changes\. \(b\) Their difference on one axis: values to the left favor more data, values to the right favor a larger model, and intervals crossing zero are unresolved\.
### 5\.3Most Tasks Favor Data at Matched Compute

The previous analysis shows that pretraining data and model size work together\. In practice, however, the training budget is fixed\. We therefore ask which direction is associated with a larger downstream gain at the same estimated compute: using more pretraining data or using a larger model\.

Figure[4](https://arxiv.org/html/2609.27232#S5.F4)\(a\) compares these two directions for each task\. Moving to the right means that performance improves more as pretraining data increase; moving upward means that performance improves more as model size increases\. Tasks below the diagonal show a larger gain from data scaling, whereas tasks above it show a larger gain from model scaling\. Panel \(b\) places the same comparison on a single axis\. Values to the left of zero favor more data, values to the right favor a larger model, and an interval crossing zero means that the current experiments cannot distinguish the two\.

Eight of the twelve tasks favor more pretraining data, while the remaining four are unresolved; none clearly favors a larger model\. This does not mean that model scaling never helps\. Rather, within the evaluated compute range, data scaling provides the more consistent improvement across downstream tasks, whereas the benefit of a larger model is more task dependent\. These comparisons describe the observed grid rather than a universal allocation rule\. The supplementary material provides the matched\-compute construction and uncertainty analysis\. Section[6](https://arxiv.org/html/2609.27232#S6)next asks whether these measurements can support a budget\-constrained allocation rule calibrated on ID tasks, and whether models selected by that rule remain competitive on OOD tasks\.

## 6OOD Evaluation of ID\-Guided Models

Section[5](https://arxiv.org/html/2609.27232#S5)shows that additional compute generally helps, but does not determine how that compute should be divided among pretraining data, model size, and training duration\. We now turn this observation into a model\-selection procedure\. Given a compute budget, we use ID downstream performance to choose a supported combination of pretraining data, model size, and training duration\. We then lock the selected model and evaluate it on OOD tasks\.

##### ID\-guided model selection\.

Following[Muennighoff et al\. \(2023\)](https://arxiv.org/html/2609.27232#bib.bib12), we fit an effective\-data response to the mean linear\-probing performance on five ID tasks\. The response relates ID performance to pretraining data size, model size, and training duration; together with the EFLOPs of each observed model, it ranks configurations only within the measured range\. Its form and parameters are selected using complete\-run ID validation and frozen before OOD evaluation\. The equations, formula comparisons, and diagnostic results are reported in the supplementary material, together with an audit of all budget\-eligible NeuroJEPA models and data\-heavy, model\-heavy, longer\-training, and best\-observed\-ID rules using OOD rank and regret\.

We refit this response using all eligible ID downstream results and consider two prespecified compute budgets that cover two baseline\-compute regimes\. At each budget, we select the observed combination of pretraining data, model size, and training duration with the best predicted ID performance\. The two selected models and all evaluation settings are locked before OOD evaluation\. The supplementary material reports validation that holds out complete training runs or one data scale at a time\.

The OOD datasets also appear in the descriptive scaling analyses in Section[5](https://arxiv.org/html/2609.27232#S5)\. However, neither their validation nor test performance enters the response fit, compute\-budget definition, or model selection in this section, and the OOD datasets are excluded from NeuroJEPA pretraining\. This is therefore a selection\-held\-out OOD evaluation, not a claim that these datasets were absent from every descriptive analysis\. The exact fitting and compute\-matching procedures are given in the supplementary material\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/fig5.jpg)Figure 5:OOD performance versus estimated pretraining compute\.Blue markers denote the two NeuroJEPA models selected using ID performance only, and red markers denote public fMRI foundation\-model baselines; labels report model parameters\. The vertical axis averages classification accuracy and Pearson correlation for regression across the OOD tasks in Table[2](https://arxiv.org/html/2609.27232#S6.T2)\.ModelModelparametersPretrainEFLOPsADNI\-ADADHDNKI\-AgeAcc↑\\uparrowF1↑\\uparrowAcc↑\\uparrowF1↑\\uparrowMSE↓\\downarrowr↑r\\uparrowBrainLM\-13M13M0\.83556\.4±\\pm3\.656\.4±\\pm3\.658\.9±\\pm4\.057\.4±\\pm4\.7\.937±\\pm\.027\.321±\\pm\.015BrainLM\-111M111M206\.7776\.4±\\pm3\.076\.4±\\pm3\.052\.1±\\pm2\.850\.4±\\pm1\.1\.776±\\pm\.056\.503±\\pm\.039BrainMASS14\.4M2\.75560\.4±\\pm4\.160\.0±\\pm4\.055\.9±\\pm0\.955\.7±\\pm1\.1\.816±\\pm\.054\.460±\\pm\.047BrainGFM10\.3M0\.12960\.9±\\pm3\.060\.7±\\pm3\.155\.4±\\pm1\.652\.7±\\pm3\.0\.981±\\pm\.026\.248±\\pm\.043BrainHarmonix\-F85\.25M110\.9069\.8±\\pm5\.068\.8±\\pm5\.553\.4±\\pm1\.147\.0±\\pm8\.3\.878±\\pm\.066\.394±\\pm\.048Brain\-Semantoks129\.21M0\.175567\.1±\\pm6\.567\.1±\\pm6\.558\.3±\\pm1\.9457\.4±\\pm1\.94\.741±\\pm\.040\.533±\\pm\.031NeuroJEPA2\.65M0\.12871\.1±\\pm4\.270\.8±\\pm4\.554\.52±\\pm2\.050\.78±\\pm7\.4\.573±\\pm\.028\.669±\\pm\.034NeuroJEPA10\.42M1\.9767\.1±\\pm3\.866\.7±\\pm3\.360\.3±\\pm1\.060\.1±\\pm1\.2\.438±\\pm\.029\.761±\\pm\.023
ModelModelparametersPretrainEFLOPsBHRC\-SexNKI\-Edu\.SALD\-AgeAcc↑\\uparrowF1↑\\uparrowAcc↑\\uparrowF1↑\\uparrowMSE↓\\downarrowr↑r\\uparrowBrainLM\-13M13M0\.83558\.0±\\pm3\.848\.4±\\pm8\.952\.5±\\pm3\.746\.7±\\pm4\.3\.915±\\pm\.122\.382±\\pm\.071BrainLM\-111M111M206\.7756\.0±\\pm5\.944\.2±\\pm3\.948\.7±\\pm4\.046\.2±\\pm5\.3\.725±\\pm\.044\.548±\\pm\.051BrainMASS14\.4M2\.75561\.1±\\pm5\.246\.8±\\pm6\.456\.2±\\pm3\.054\.9±\\pm4\.6\.781±\\pm\.057\.476±\\pm\.036BrainGFM10\.3M0\.12956\.6±\\pm5\.548\.9±\\pm3\.432\.8±\\pm4\.726\.2±\\pm5\.41\.067±\\pm\.067\-\.021±\\pm\.054BrainHarmonix\-F85\.25M110\.9059\.4±\\pm4\.349\.2±\\pm10\.138\.1±\\pm4\.536\.9±\\pm4\.7\.824±\\pm\.100\.422±\\pm\.081Brain\-Semantoks129\.21M0\.175558\.9±\\pm4\.053\.2±\\pm9\.653\.6±\\pm8\.052\.0±\\pm7\.9\.557±\\pm\.037\.681±\\pm\.016NeuroJEPA2\.65M0\.12857\.7±\\pm3\.345\.9±\\pm6\.1957\.7±\\pm5\.355\.8±\\pm5\.3\.505±\\pm\.035\.706±\\pm\.016NeuroJEPA10\.42M1\.9758\.9±\\pm3\.156\.2±\\pm8\.062\.6±\\pm4\.8362\.2±\\pm4\.86\.428±\\pm\.020\.762±\\pm\.028

Table 2:Downstream evaluation with frozen linear probes\. Classification tasks report accuracy and F1 in percent; Regression tasks report MSE and Pearson correlation\. Each entry is the mean±\\pmstandard deviation across seeds\. Best and second\-best means are shown in bold and underlined, respectively\. ADNI\-MCI results are reported in the supplementary material\.
##### Evaluation protocol\.

We evaluate the six OOD tasks reported in Table[2](https://arxiv.org/html/2609.27232#S6.T2): AD classification, ADHD diagnosis, BHRC sex classification, SALD age prediction, NKI age prediction, and NKI education classification\. Every encoder is frozen, and the same linear\-probe search is applied to all models\. Classification tasks useL2L\_\{2\}\-regularized logistic regression, whereas regression tasks use ridge regression\. Input features are standardized using statistics computed from the training split\. For regression, targets are also standardized using the training\-set mean and standard deviation\. The regularization strength is selected separately for every model, task, and run on the validation split, using log loss for classification and mean\-squared error for regression\. The selected probe, trained only on the training partition, is then evaluated once on the corresponding test partition\.

All splits are made at the subject level, with every recording from one participant kept in the same split\. The training partition contains 60% of the subjects and remains fixed\. The original validation and test subjects are combined into a 40% holdout pool\. Before each of five runs, this pool is randomly divided in half using a different seed, producing a 60:20:20 train–validation–test split\. The same seed\-specific split is used for every baseline and both NeuroJEPA configurations\.

##### Baseline and compute accounting\.

Each fMRI foundation\-model baseline is evaluated from its official released weights using the accompanying code and required input preprocessing\. For every baseline and both selected NeuroJEPA models, we report the number of trainable model parameters and estimated pretraining EFLOPs\. Baseline EFLOPs are reconstructed from the training details reported in the paper, released code, configuration files, and checkpoint metadata\. Because these values are estimates, we report their assumptions and available ranges in the supplementary material\.

##### OOD results\.

Figure[5](https://arxiv.org/html/2609.27232#S6.F5)shows that both ID\-guided NeuroJEPA models achieve a favorable OOD performance–compute trade\-off, with the 10\.42M model obtaining the highest average OOD performance among the compared models\. Table[2](https://arxiv.org/html/2609.27232#S6.T2)shows that this result is not driven by a single dataset: the 10\.42M model achieves the best mean on at least one reported metric in five of the six tasks, while the 2\.65M model ranks second on at least one metric in four tasks\. Relative to the closest\-compute baseline, each selected model obtains a higher mean on five of the six task\-wise metrics summarized in Figure[5](https://arxiv.org/html/2609.27232#S6.F5)\. These results support ID\-guided allocation as a compute\-efficient model\-selection strategy, without implying that one configuration is optimal for every OOD task\.

## 7Discussion and Limitations

Estimated compute is a useful measure of training cost, but it does not uniquely determine downstream performance\. Performance generally improves as training advances, yet configurations with similar EFLOPs can remain separated\. The fitted data\-size association becomes more favorable as model size increases and is stronger at the end of training\. At matched compute, it is larger than the model\-size association for eight of twelve tasks; the other four are unresolved\. The configuration behind a budget therefore matters in addition to its total compute\.

The practical question is which combination of pretraining data size, model size, and training progress makes the best use of fixed compute\. Our results motivate searching these choices jointly rather than treating a larger model as a substitute for more data or training\. Data\-rich configurations are important candidates within the evaluated range, but the results do not prescribe a universal data–model ratio\.

We make this decision concrete by fitting ID downstream performance over the supported combinations\. At two compute budgets, we select the combination with the best predicted ID performance and lock it before OOD evaluation\. The selected models remain competitive across six OOD tasks, and the larger one achieves the highest average OOD performance among the compared fMRI foundation models\. Thus, ID performance can guide a strong fixed\-budget choice, without showing that the selected combination is optimal for every OOD task or beyond the evaluated candidates\.

The main limitation concerns extrapolation of the fitted ID response, not the relationships observed within the current grid\. The six nested data sizes span approximately 24\.5\-fold, with no substantially larger configuration reserved for a prospective test\. The response can compare configurations within or close to this range but cannot predict far beyond it\. Holding source datasets and sampling proportions fixed strengthens the comparison; a future corpus with new sites, scanners, or acquisition protocols may require the numerical response to be fitted again\.

Finally, we study one fixed method and ROI\-time architecture family with frozen linear probes\. Task resampling omits independent pretraining\-run variation, and baseline EFLOPs are estimated\. These limitations motivate larger held\-out scales and replicated key runs\.

## References

- Balestriero and LeCun \(2025\)Randall Balestriero and Yann LeCun\.LeJEPA: Provable and scalable self\-supervised learning without the heuristics\.*arXiv preprint arXiv:2511\.08544*, 2025\.doi:10\.48550/arXiv\.2511\.08544\.URL[https://arxiv\.org/abs/2511\.08544](https://arxiv.org/abs/2511.08544)\.
- Cherti et al\. \(2023\)Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev\.Reproducible scaling laws for contrastive language\-image learning\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, pages 2818–2829, 2023\.URL[https://openaccess\.thecvf\.com/content/CVPR2023/html/Cherti\_Reproducible\_Scaling\_Laws\_for\_Contrastive\_Language\-Image\_Learning\_CVPR\_2023\_paper\.html](https://openaccess.thecvf.com/content/CVPR2023/html/Cherti_Reproducible_Scaling_Laws_for_Contrastive_Language-Image_Learning_CVPR_2023_paper.html)\.
- Di Martino et al\. \(2014\)Adriana Di Martino, Chao\-Gan Yan, Qingyang Li, et al\.The autism brain imaging data exchange: towards a large\-scale evaluation of the intrinsic brain architecture in autism\.*Molecular Psychiatry*, 19\(6\):659–667, 2014\.doi:10\.1038/mp\.2013\.78\.
- Dong et al\. \(2025\)Zijian Dong, Ruilin Li, Joanna Su Xian Chong, Niousha Dehestani, Yinghui Teng, Yi Lin, Zhizhou Li, Yichi Zhang, Yapei Xie, Leon Qi Rong Ooi, B\. T\. Thomas Yeo, and Juan Helen Zhou\.Brain harmony: A multimodal foundation model unifying morphology and function into 1d tokens\.*arXiv preprint arXiv:2509\.24693*, 2025\.doi:10\.48550/arXiv\.2509\.24693\.URL[https://arxiv\.org/abs/2509\.24693](https://arxiv.org/abs/2509.24693)\.NeurIPS 2025\.
- Gijsen et al\. \(2026\)Sam Gijsen, Marc\-Andre Schulz, and Kerstin Ritter\.Brain\-Semantoks: Learning semantic tokens of brain dynamics with a self\-distilled foundation model\.In*International Conference on Learning Representations*, 2026\.URL[https://iclr\.cc/virtual/2026/poster/10011042](https://iclr.cc/virtual/2026/poster/10011042)\.
- Hoffmann et al\. \(2022\)Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W\. Rae, Oriol Vinyals, and Laurent Sifre\.Training compute\-optimal large language models\.In*Advances in Neural Information Processing Systems*, volume 35, pages 30016–30030, 2022\.URL[https://arxiv\.org/abs/2203\.15556](https://arxiv.org/abs/2203.15556)\.
- Jack et al\. \(2008\)Clifford R\. Jack, Matt A\. Bernstein, Nick C\. Fox, et al\.The alzheimer’s disease neuroimaging initiative \(ADNI\): MRI methods\.*Journal of Magnetic Resonance Imaging*, 27\(4\):685–691, 2008\.doi:10\.1002/jmri\.21049\.
- Kaplan et al\. \(2020\)Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B\. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei\.Scaling laws for neural language models\.*arXiv preprint arXiv:2001\.08361*, 2020\.doi:10\.48550/arXiv\.2001\.08361\.URL[https://arxiv\.org/abs/2001\.08361](https://arxiv.org/abs/2001.08361)\.
- Lane et al\. \(2025\)Connor Lane, Mihir Tripathy, Leema Krishna Murali, Ratna Sagari Grandhi, Shamus Sim Zi Yang, Sam Gijsen, Debojyoti Das, Manish Ram, Utkarsh Kumar Singh, Cesar Kadir Torrico Villanueva, Yuxiang Wei, Will Beddow, Gianfranco Cortés, Suin Cho, Daniel Z\. Kaplan, Benjamin Warner, Tanishq Mathew Abraham, and Paul S\. Scotti\.Scaling vision transformers for functional MRI with flat maps\.*arXiv preprint arXiv:2510\.13768*, 2025\.doi:10\.48550/arXiv\.2510\.13768\.URL[https://arxiv\.org/abs/2510\.13768](https://arxiv.org/abs/2510.13768)\.Accepted at ICML 2026\.
- Marek et al\. \(2011\)Kenneth Marek, Danna Jennings, Shirley Lasch, et al\.The parkinson progression marker initiative \(PPMI\)\.*Progress in Neurobiology*, 95\(4\):629–635, 2011\.doi:10\.1016/j\.pneurobio\.2011\.09\.005\.
- Marek et al\. \(2022\)Scott Marek, Brenden Tervo\-Clemmens, Finnegan J\. Calabro, David F\. Montez, Benjamin P\. Kay, Alexander S\. Hatoum, Meghan Rose Donohue, William Foran, Ryland L\. Miller, Timothy J\. Hendrickson, Stephen M\. Malone, Sridhar Kandala, Eric Feczko, Oscar Miranda\-Dominguez, Alice M\. Graham, Eric A\. Earl, Anders J\. Perrone, Michaela Cordova, Olivia Doyle, Lucille A\. Moore, Gregory M\. Conan, Johnny Uriarte, Kathy Snider, Benjamin J\. Lynch, James C\. Wilgenbusch, Thomas Pengo, Angela Tam, Jianzhong Chen, Dillan J\. Newbold, Annie Zheng, Nicole A\. Seider, Andrew N\. Van, Athanasia Metoki, Roselyne J\. Chauvin, Timothy O\. Laumann, Deanna J\. Greene, Steven E\. Petersen, Hugh Garavan, Wesley K\. Thompson, Thomas E\. Nichols, B\. T\. Thomas Yeo, Deanna M\. Barch, Beatriz Luna, Damien A\. Fair, and Nico U\. F\. Dosenbach\.Reproducible brain\-wide association studies require thousands of individuals\.*Nature*, 603\(7902\):654–660, 2022\.doi:10\.1038/s41586\-022\-04492\-9\.URL[https://www\.nature\.com/articles/s41586\-022\-04492\-9](https://www.nature.com/articles/s41586-022-04492-9)\.
- Muennighoff et al\. \(2023\)Niklas Muennighoff, Alexander M\. Rush, Boaz Barak, Teven Le Scao, Aleksandra Piktus, Nouamane Tazi, Sampo Pyysalo, Thomas Wolf, and Colin Raffel\.Scaling data\-constrained language models\.*arXiv preprint arXiv:2305\.16264*, 2023\.doi:10\.48550/arXiv\.2305\.16264\.URL[https://arxiv\.org/abs/2305\.16264](https://arxiv.org/abs/2305.16264)\.
- Nooner et al\. \(2012\)Kate Brody Nooner, Stanley J\. Colcombe, Russell H\. Tobe, et al\.The NKI\-rockland sample: A model for accelerating the pace of discovery science in psychiatry\.*Frontiers in Neuroscience*, 6:152, 2012\.doi:10\.3389/fnins\.2012\.00152\.
- Ortega Caro et al\. \(2024\)Josue Ortega Caro, Antonio Henrique de Oliveira Fonseca, Syed Rizvi, Matteo Rosati, Christopher Averill, James Cross, Prateek Mittal, Emanuele Zappala, Rahul Dhodapkar, Chadi Abdallah, and David van Dijk\.BrainLM: A foundation model for brain activity recordings\.In*International Conference on Learning Representations*, pages 565–576, 2024\.URL[https://proceedings\.iclr\.cc/paper\_files/paper/2024/hash/029ce70401321de3808b3ac39e1ab167\-Abstract\-Conference\.html](https://proceedings.iclr.cc/paper_files/paper/2024/hash/029ce70401321de3808b3ac39e1ab167-Abstract-Conference.html)\.Verified against the official ICLR 2024 proceedings page\.
- Salum et al\. \(2025\)Giovanni A\. Salum, Carina de Giusti, Laila Souza, et al\.Cohort profile: Brazilian high\-risk cohort for mental health conditions \(BHRC\)\.*International Journal of Epidemiology*, 54\(6\), 2025\.doi:10\.1093/ije/dyaf192\.
- Satterthwaite et al\. \(2014\)Theodore D\. Satterthwaite, Mark A\. Elliott, Kosha Ruparel, et al\.Neuroimaging of the philadelphia neurodevelopmental cohort\.*NeuroImage*, 86:544–553, 2014\.doi:10\.1016/j\.neuroimage\.2013\.07\.064\.
- Schaefer et al\. \(2018\)Alexander Schaefer, Ru Kong, Evan M\. Gordon, Timothy O\. Laumann, Xi\-Nian Zuo, Avram J\. Holmes, Simon B\. Eickhoff, and B\. T\. Thomas Yeo\.Local\-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI\.*Cerebral Cortex*, 28\(9\):3095–3114, 2018\.doi:10\.1093/cercor/bhx179\.
- The ADHD\-200 Consortium \(2012\)The ADHD\-200 Consortium\.The ADHD\-200 consortium: a model to advance the translational potential of neuroimaging in clinical neuroscience\.*Frontiers in Systems Neuroscience*, 6:62, 2012\.doi:10\.3389/fnsys\.2012\.00062\.
- Van Essen et al\. \(2013\)David C\. Van Essen, Stephen M\. Smith, Deanna M\. Barch, Timothy E\. J\. Behrens, Essa Yacoub, and Kamil Ugurbil\.The WU\-minn human connectome project: An overview\.*NeuroImage*, 80:62–79, 2013\.doi:10\.1016/j\.neuroimage\.2013\.05\.041\.
- Wang et al\. \(2025\)Mo Wang, Junfeng Xia, Wenhao Ye, Enyu Liu, Kaining Peng, Jianfeng Feng, Quanying Liu, and Hongkai Wen\.SLIM\-Brain: A data\- and training\-efficient foundation model for fMRI data analysis\.*arXiv preprint arXiv:2512\.21881*, 2025\.doi:10\.48550/arXiv\.2512\.21881\.URL[https://arxiv\.org/abs/2512\.21881](https://arxiv.org/abs/2512.21881)\.
- Wang et al\. \(2026a\)Mo Wang, Wenhao Ye, Junfeng Xia, Minghao Xu, Hongkai Wen, and Quanying Liu\.FlexiBrain: Resolution\-agnostic voxel\-level encoding for native fMRI\.*arXiv preprint arXiv:2606\.11500*, 2026a\.doi:10\.48550/arXiv\.2606\.11500\.URL[https://arxiv\.org/abs/2606\.11500](https://arxiv.org/abs/2606.11500)\.
- Wang et al\. \(2026b\)Mo Wang, Wenhao Ye, Junfeng Xia, Junxiang Zhang, Xuanye Pan, Minghao Xu, Haotian Deng, Hongkai Wen, and Quanying Liu\.Omni\-fMRI: A universal atlas\-free fMRI foundation model\.*arXiv preprint arXiv:2601\.23090*, 2026b\.doi:10\.48550/arXiv\.2601\.23090\.URL[https://arxiv\.org/abs/2601\.23090](https://arxiv.org/abs/2601.23090)\.
- Wei et al\. \(2018\)Dongtao Wei, Kaixiang Zhuang, Lei Ai, et al\.Structural and functional brain scans from the cross\-sectional southwest university adult lifespan dataset\.*Scientific Data*, 5\(1\):180134, 2018\.doi:10\.1038/sdata\.2018\.134\.
- Wei et al\. \(2025\)Xinxu Wei, Kanhao Zhao, Yong Jiao, Lifang He, and Yu Zhang\.A brain graph foundation model: Pre\-training and prompt\-tuning across broad atlases and disorders\.*arXiv preprint arXiv:2506\.02044*, 2025\.doi:10\.48550/arXiv\.2506\.02044\.URL[https://arxiv\.org/abs/2506\.02044](https://arxiv.org/abs/2506.02044)\.
- Xia et al\. \(2026a\)Junfeng Xia, Wenhao Ye, Xuanye Pan, Xinke Shen, Mo Wang, and Quanying Liu\.Brain\-DiT: A universal multi\-state fMRI foundation model with metadata\-conditioned pretraining\.*arXiv preprint arXiv:2604\.12683*, 2026a\.doi:10\.48550/arXiv\.2604\.12683\.URL[https://arxiv\.org/abs/2604\.12683](https://arxiv.org/abs/2604.12683)\.
- Xia et al\. \(2026b\)Junfeng Xia, Wenhao Ye, Junxiang Zhang, Xuanye Pan, Mo Wang, and Quanying Liu\.BrainWorld: A structural\-prior\-conditioned generative model for whole\-brain 4d fMRI dynamics\.*arXiv preprint arXiv:2606\.17742*, 2026b\.doi:10\.48550/arXiv\.2606\.17742\.URL[https://arxiv\.org/abs/2606\.17742](https://arxiv.org/abs/2606.17742)\.
- Xia et al\. \(2026c\)Junfeng Xia, Wenhao Ye, Junxiang Zhang, Jiayu Zuo, Mo Wang, and Quanying Liu\.BrainTaskonomy: Learning how to pretrain and what to transfer in fMRI foundation models\.*arXiv preprint arXiv:2609\.10518*, 2026c\.doi:10\.48550/arXiv\.2609\.10518\.URL[https://arxiv\.org/abs/2609\.10518](https://arxiv.org/abs/2609.10518)\.
- Yang et al\. \(2024\)Yanwu Yang, Chenfei Ye, Guinan Su, Ziyao Zhang, Zhikai Chang, Hairui Chen, Piu Chan, Yue Yu, and Ting Ma\.BrainMass: Advancing brain network analysis for diagnosis with large\-scale self\-supervised learning\.*IEEE Transactions on Medical Imaging*, 43\(11\):4004–4016, 2024\.doi:10\.1109/TMI\.2024\.3414476\.URL[https://doi\.org/10\.1109/TMI\.2024\.3414476](https://doi.org/10.1109/TMI.2024.3414476)\.
- Zhai et al\. \(2022\)Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer\.Scaling vision transformers\.In*Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, pages 12104–12113, 2022\.URL[https://openaccess\.thecvf\.com/content/CVPR2022/html/Zhai\_Scaling\_Vision\_Transformers\_CVPR\_2022\_paper\.html](https://openaccess.thecvf.com/content/CVPR2022/html/Zhai_Scaling_Vision_Transformers_CVPR_2022_paper.html)\.

Supplementary Material for

A Scaling Study for fMRI Foundation Models

## Appendix ASupplementary Methods and Analyses

This appendix follows the order of the main paper\. It first records the fixed pretraining and evaluation details needed for reproduction, then gives the statistical definitions behind the scaling results, and finally documents the ID\-guided model\-selection and OOD audits\.

### A\.1Fixed Pretraining Method

#### Latent Geometry and SIGReg

The geometry argument provides intuition for the fixed objective; it is not a new theorem and is not used to establish the scaling results\. LetZ∈ℝn×dembZ\\in\\mathbb\{R\}^\{n\\times d\_\{\\mathrm\{emb\}\}\}containnncentered frozen embeddings and letμ1,…,μr\\mu\_\{1\},\\ldots,\\mu\_\{r\}be the positive eigenvalues ofΣ^=Z⊤​Z/n\\widehat\{\\Sigma\}=Z^\{\\top\}Z/n\. At fixed rank and total variance, distributing variance evenly avoids directions with very little variation\. In an idealized linear model, this has two familiar consequences\. For ridge regression with penaltyλr\\lambda\_\{\\mathrm\{r\}\}and a target vector of normρ\\rho, the largest directional shrinkage is controlled by the smallest positive eigenvalue,

sup‖β⋆‖2=ρ‖𝔼⁡\[β^∣Z\]−β⋆‖22=λr2​ρ2\(μmin\+λr\)2\.\\sup\_\{\\\|\\beta^\{\\star\}\\\|\_\{2\}=\\rho\}\\left\\\|\\mathbb\{E\}\[\\widehat\{\\beta\}\\mid Z\]\-\\beta^\{\\star\}\\right\\\|\_\{2\}^\{2\}=\\frac\{\\lambda\_\{\\mathrm\{r\}\}^\{2\}\\rho^\{2\}\}\{\(\\mu\_\{\\min\}\+\\lambda\_\{\\mathrm\{r\}\}\)^\{2\}\}\.Under homoscedastic noise with varianceσ2\\sigma^\{2\}, the total ordinary least\-squares coefficient variance on the rank\-rrrepresentation subspace is proportional to

σ2n​∑j=1r1μj\.\\frac\{\\sigma^\{2\}\}\{n\}\\sum\_\{j=1\}^\{r\}\\frac\{1\}\{\\mu\_\{j\}\}\.Equal positive eigenvalues minimize this quantity when their sum is fixed\. These calculations explain why a well\-spread representation can be convenient for unknown linear readouts and label\-limited probes\. They do not imply that isotropy creates task information, improves every downstream task, or makes the objective optimal for fMRI\.

In NeuroJEPA, alignment encourages global and local views of the same ROI\-time segment to share a representation\. SIGReg separately discourages collapse and concentration into a few latent directions\. We use the Epps–Pulley SIGReg functional from LeJEPA\([Balestriero and LeCun, 2025](https://arxiv.org/html/2609.27232#bib.bib1)\)unchanged\. For a batch of embeddings𝒵=\{zi\}i=1B\\mathcal\{Z\}=\\\{z\_\{i\}\\\}\_\{i=1\}^\{B\}andMMrandom unit directions𝒜=\{am\}m=1M\\mathcal\{A\}=\\\{a\_\{m\}\\\}\_\{m=1\}^\{M\}inℝdz\\mathbb\{R\}^\{d\_\{z\}\},

ℒSIGReg​\(𝒵\)=1M​∑m=1MTEP​\(\{am⊤​zi\}i=1B\),\\mathcal\{L\}\_\{\\mathrm\{SIGReg\}\}\(\\mathcal\{Z\}\)=\\frac\{1\}\{M\}\\sum\_\{m=1\}^\{M\}T\_\{\\mathrm\{EP\}\}\\\!\\left\(\\\{a\_\{m\}^\{\\top\}z\_\{i\}\\\}\_\{i=1\}^\{B\}\\right\),whereTEPT\_\{\\mathrm\{EP\}\}compares the projected empirical distribution with a standard Gaussian\. We useM=4096M=4096directions and 17 trapezoidal integration points on\[−5,5\]\[\-5,5\]\. Directions are sampled from a standard Gaussian, normalized to unit length, and synchronized across workers\. Empirical characteristic functions are averaged across workers before the statistic is evaluated\. SIGReg is computed separately for each of the eight view batches and then averaged\.

#### View Construction

Letx~i∈ℝR×Tiraw\\widetilde\{x\}\_\{i\}\\in\\mathbb\{R\}^\{R\\times T\_\{i\}^\{\\mathrm\{raw\}\}\}denote one recording\. We first sample the segmentxi∈ℝR×Tx\_\{i\}\\in\\mathbb\{R\}^\{R\\times T\}used by the objective in Section[3](https://arxiv.org/html/2609.27232#S3)\. Two global views retain broad context and six local views retain less context\. Cropped views are resampled to fixed shapes, perturbed, and independently z\-scored over time within each ROI\. Table[3](https://arxiv.org/html/2609.27232#A1.T3)gives the complete recipe\.

Table 3:View construction and perturbation recipe\.Shifts are circular and measured in time points\. Before view construction, temporal resampling is applied with probability0\.20\.2using a scale sampled from\[0\.9,1\.2\]\[0\.9,1\.2\]and linear interpolation\.
#### Architecture and Optimization

Table[4](https://arxiv.org/html/2609.27232#A1.T4)lists settings shared by all scaling runs\. View\-specific settings are reported once in Table[3](https://arxiv.org/html/2609.27232#A1.T3)\.

#### NeuroJEPA Compute Accounting

We count the dominant dense matrix multiplications in the encoder and projection head\. One multiply–accumulate is two floating\-point operations\. Normalization, activation functions, positional encoding, and objective statistics are excluded\. For viewvvwith spatial dimensions\(Hv,Wv\)\(H\_\{v\},W\_\{v\}\), patch size\(ph,pw\)\(p\_\{h\},p\_\{w\}\), embedding dimensiondd, andnregn\_\{\\mathrm\{reg\}\}register tokens, define

Pv=⌊Hvph⌋​⌊Wvpw⌋,Tv=Pv\+1\+nreg\.P\_\{v\}=\\left\\lfloor\\frac\{H\_\{v\}\}\{p\_\{h\}\}\\right\\rfloor\\left\\lfloor\\frac\{W\_\{v\}\}\{p\_\{w\}\}\\right\\rfloor,\\qquad T\_\{v\}=P\_\{v\}\+1\+n\_\{\\mathrm\{reg\}\}\.HerePvP\_\{v\}is the number of patch tokens andTvT\_\{v\}also includes the class and register tokens\. With patch areaA=ph​pwA=p\_\{h\}p\_\{w\}, Transformer depthLL, and MLP ratiorr, the forward cost of one view is

Fview​\(v\)=2​Pv​A​d\+L⁡\[\(8\+4​r\)​Tv​d2\+4​Tv2​d\]\.F\_\{\\mathrm\{view\}\}\(v\)=2P\_\{v\}Ad\+L\\\!\\left\[\(8\+4r\)T\_\{v\}d^\{2\}\+4T\_\{v\}^\{2\}d\\right\]\.For a projection headd→Hproj→Hproj→dzd\\rightarrow H\_\{\\mathrm\{proj\}\}\\rightarrow H\_\{\\mathrm\{proj\}\}\\rightarrow d\_\{z\},

Fproj=\(d​Hproj\+Hproj2\+Hproj​dz\)\.F\_\{\\mathrm\{proj\}\}=2\\\!\\left\(dH\_\{\\mathrm\{proj\}\}\+H\_\{\\mathrm\{proj\}\}^\{2\}\+H\_\{\\mathrm\{proj\}\}d\_\{z\}\\right\)\.Withng=2n\_\{g\}=2global views andnl=6n\_\{l\}=6local views, we approximate backward computation as twice the forward cost:

Ftrain/sample=3\[\\displaystyle F\_\{\\mathrm\{train/sample\}\}=3\\big\[ng​Fview​\(g\)\+nl​Fview​\(l\)\\displaystyle n\_\{g\}F\_\{\\mathrm\{view\}\}\(g\)\+n\_\{l\}F\_\{\\mathrm\{view\}\}\(l\)\+\(ng\+nl\)Fproj\]\.\\displaystyle\+\(n\_\{g\}\+n\_\{l\}\)F\_\{\\mathrm\{proj\}\}\\big\]\.LetKKbe the number of completed optimizer steps andBglobalB\_\{\\mathrm\{global\}\}the effective batch size\. Cumulative compute is

C⁡\(K\)=K​Bglobal​Ftrain/sample1018​EFLOP\.\\boxed\{C\(K\)=\\frac\{KB\_\{\\mathrm\{global\}\}F\_\{\\mathrm\{train/sample\}\}\}\{10^\{18\}\}\\ \\text\{EFLOP\}\.\}Thus, equal parameter counts need not imply equal compute: view shapes, batch size, and the number of completed steps also enter the estimate\.

ParameterMeaningValueEncoder and projectorInput segmentSchaefer\-100 cortical ROI\-time matrix100×200100\\times 200Patch sizeROI\-by\-time patch size1×401\\times 40Summary / register tokensLearned non\-patch tokens1/41/4Encoder width and depthConfiguration\-specific valuesTable[5](https://arxiv.org/html/2609.27232#A1.T5)Transformer MLP ratioHidden\-to\-embedding width ratio44Projection headShared MLP after the encoderdemb→→→512d\_\{\\mathrm\{emb\}\}\\\!\\rightarrow\\\!2048\\\!\\rightarrow\\\!2048\\\!\\rightarrow\\\!512Attention normalizationQKV bias and query/key normalizationBias enabled; LayerNormResidual regularizationDrop path and LayerScale initialization0\.10\.1;10−410^\{\-4\}Position encodingLearned positions plus RoPErotary ratio0\.50\.5, base300300ObjectiveLoss weightsAlignment / SIGReg0\.95/0\.050\.95/0\.05SIGReg settingsHyperparameter, projections, integration pointsαSIG=0\.5\\alpha\_\{\\mathrm\{SIG\}\}=0\.5;40964096;1717Regularized dimensionProjection dimension seen by SIGReg512512OptimizationOptimizerDecoupled weight decayAdamW,\(0\.9,0\.999\)\(0\.9,0\.999\)Learning rateReference rate and batch scaling0\.01​Bglobal/10240\.01\\sqrt\{B\_\{\\mathrm\{global\}\}/1024\}ScheduleWarm\-up, cosine decay, terminal floor10%10\\%; cosine;10−610^\{\-6\}Weight decayInitial value5×10−45\\times 10^\{\-4\}Layer\-wise LR decayMultiplicative decay across depth0\.90\.9Gradient clipping / accumulationGlobal norm; accumulation steps3\.0/13\.0/1Table 4:Shared implementation settings\.The global batch size and number of optimizer steps are run specific\.Table 5:Encoder configurations used in the scaling experiments\.Parameter counts exclude the projection head\.
#### Experimental Resource Use

The experiments used approximately 10,347 GPU\-hours in total\. This quantity documents the computational resources consumed by the study; it is distinct from the EFLOP estimates used for model comparison, which account for each model’s computation rather than hardware occupancy\. Pretraining accounts for most of the recorded GPU\-hours, and most experiments were run on NVIDIA A800\-SXM4\-80GB GPUs\. All values below are rounded\.

Table 6:GPU\-hours by experiment stage\.

Table 7:GPU\-hours by hardware type\.
Pretraining recordingsMerged GPU\-hoursDownstream allocationFinal GPU\-hoursShare63,4014,766\.7723\.884,790\.6546\.30%32,414672\.0214\.45686\.476\.63%11,673995\.640\.12995\.769\.62%6,5331,275\.540\.361,275\.9012\.33%3,810984\.220\.24984\.469\.51%2,5851,436\.648\.561,445\.2013\.97%Six\-scale subtotal10,130\.8247\.6110,178\.4398\.37%All\-experiment total10,299\.7247\.6110,347\.33100%
Table 8:GPU\-hours attributed to each pretraining\-data scale\.The downstream\-allocation column records the data\-scale\-specific downstream cost assigned in the experiment ledger\. Rows show the six nested data sizes used in the main scaling analysis\. The subtotal covers those six scales, whereas the all\-experiment total matches the stage and hardware totals above and also includes experiments outside the main six\-scale analysis\. Small discrepancies between displayed row sums and totals are due to rounding\.

### A\.2Data and Downstream Evaluation

The full source catalogue, including participant and recording counts, appears in Appendix[C](https://arxiv.org/html/2609.27232#A3)\. The six pretraining data sizes in the main analysis are subject\-grouped, strictly nested subsets of that catalogue\. Downstream validation and test participants are excluded from pretraining\. For the final evaluation in Section[6](https://arxiv.org/html/2609.27232#S6), the OOD datasets are also excluded from the selected models’ pretraining data\. Their downstream scores are used in the descriptive scaling analyses, but only the five ID tasks are used to fit the response and select the two final models\.

Classification summaries use macro\-F1 and regression summaries use Pearson correlation\. Both are oriented so that higher is better\. We first average each task over five probe seeds and then give every task equal weight\. Accuracy and mean\-squared error are reported as additional task\-level metrics but do not enter these cross\-task summaries\. In the interaction analyses below, task\-wise standardization is applied only to place regression coefficients from different tasks on a common numerical scale\.

#### Frozen Linear Probes

For each frozen model, task, and evaluation seed, the feature standardizer is fit only on the training partition\. Classification usesL2L\_\{2\}\-regularized logistic regression with inverse regularization strength

κlogit∈\{10−5,10−4,…,103\},\\kappa\_\{\\mathrm\{logit\}\}\\in\\\{10^\{\-5\},10^\{\-4\},\\ldots,10^\{3\}\\\},and regression uses ridge regression with

λridge∈\{105,104,…,10−4\}\.\\lambda\_\{\\mathrm\{ridge\}\}\\in\\\{10^\{5\},10^\{4\},\\ldots,10^\{\-4\}\\\}\.Regression targets are standardized using the training\-set mean and standard deviation\. Validation log loss selectsκlogit\\kappa\_\{\\mathrm\{logit\}\}, and validation mean\-squared error selectsλridge\\lambda\_\{\\mathrm\{ridge\}\}\. Logistic regression is optimized for at most 5,000 iterations using the run seed\. Exact ties follow the fixed grid order, which favors stronger regularization\. The training\-fitted probe is evaluated directly on the test partition and is not refit on the combined training and validation data\.

Table 9:Downstream task definitions\.Participant counts and reported metrics are given in Table[1](https://arxiv.org/html/2609.27232#S4.T1)\.

### A\.3Evidence for the Scaling Findings

Figures[6](https://arxiv.org/html/2609.27232#A1.F6)and[7](https://arxiv.org/html/2609.27232#A1.F7)expand the compute trajectories in Figure[3](https://arxiv.org/html/2609.27232#S5.F3)\(a\) along the two resource axes\. The first groups models by pretraining data size, while the second groups them by model size\.

#### How Data and Model Size Interact During Training

This analysis asks whether the fitted gain associated with additional pretraining data changes with model size\. Letiiindex an observed data–model configuration,tta downstream task, andppa measured fraction of the configuration’s planned training schedule\. We use the validation score for each downstream task; the same validation split is also used to choose the probe regularization described above\. The score is oriented so that higher is better and standardized within task\. Letd~i\\widetilde\{d\}\_\{i\}andn~i\\widetilde\{n\}\_\{i\}denote standardized log data size and standardized log model size\. Their standardization is held fixed acrosspp\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/appendix_fig1.jpg)Figure 6:Training trajectories grouped by pretraining data size\.Each panel fixes the number of pretraining recordings and compares model sizes as corrected pretraining compute increases\. Lower linear\-probing error is better\.For every task and training fraction, we fit

si,t​\(p\)=\\displaystyle s\_\{i,t\}\(p\)=\{\}at​\(p\)\+bD,t​\(p\)​d~i\+bN,t​\(p\)​n~i\\displaystyle a\_\{t\}\(p\)\+b\_\{D,t\}\(p\)\\widetilde\{d\}\_\{i\}\+b\_\{N,t\}\(p\)\\widetilde\{n\}\_\{i\}\+βD×N,t​\(p\)​d~i​n~i\+ϵi,t​\(p\)\.\\displaystyle\+\\beta\_\{D\\times N,t\}\(p\)\\widetilde\{d\}\_\{i\}\\widetilde\{n\}\_\{i\}\+\\epsilon\_\{i,t\}\(p\)\.The coefficientβD×N,t​\(p\)\\beta\_\{D\\times N,t\}\(p\)answers the question directly: a positive value means that the fitted data\-size gain is larger for larger models, or equivalently that the fitted model\-size gain is larger at greater data size\. We give every task equal weight and define

βD×N​\(p\)\\displaystyle\\beta\_\{D\\times N\}\(p\)=112​∑t=112βD×N,t​\(p\),\\displaystyle=\\frac\{1\}\{12\}\\sum\_\{t=1\}^\{12\}\\beta\_\{D\\times N,t\}\(p\),Δ​β\\displaystyle\\Delta\\beta=βD×N​\(1\)−βD×N​\(0\.5\)\.\\displaystyle=\\beta\_\{D\\times N\}\(1\)\-\\beta\_\{D\\times N\}\(0\.5\)\.
Table 10:Data–model\-size coefficient across training\.The endpoint\-minus\-midpoint contrast isΔ​β=0\.116\\Delta\\beta=0\.116, with paired task\-bootstrap 95% interval\[0\.006,0\.216\]\[0\.006,0\.216\]\.
The probability of a strictly increasing ordering across all four stages is 0\.645\. We therefore interpret the result as broad strengthening from mid to late training, not as a strictly monotonic stage\-by\-stage increase\. The paired bootstrap resamples the 12 tasks as intact units and uses the same resampled set at everypp\. Its interval measures sensitivity to the composition of the fixed task panel; it does not represent variation across independent pretraining runs or untrained configurations\.

![Refer to caption](https://arxiv.org/html/2609.27232v1/figures/appendix_fig2.jpg)Figure 7:Training trajectories grouped by model size\.Each panel fixes model size and compares pretraining data sizes as corrected pretraining compute increases\. Lower linear\-probing error is better\.
#### Data versus Model Size at Matched Compute

This analysis asks which resource direction has the larger fitted relationship with downstream performance when estimated compute is held fixed\. We use 12 logarithmically spaced slices from 0\.11 to 2\.60 EFLOP\. A trajectory contributes to slicecconly when two saved models bracket that compute value inlog10⁡C\\log\_\{10\}C; values are linearly interpolated between those models and never extrapolated\. Each retained slice contains at least four model sizes, with at least five data sizes represented for every included model size\. The supported cell count decreases from 34–35 at lower compute to 23 at the two highest slices because fewer trajectories bracket those budgets\.

At each task and compute slice, we fit

si,t​\(c\)=\\displaystyle s\_\{i,t\}\(c\)=\{\}at​\(c\)\+βD,t​\(c\)​d~i​\(c\)\+βN,t​\(c\)​n~i​\(c\)\\displaystyle a\_\{t\}\(c\)\+\\beta\_\{D,t\}\(c\)\\widetilde\{d\}\_\{i\}\(c\)\+\\beta\_\{N,t\}\(c\)\\widetilde\{n\}\_\{i\}\(c\)\+ηt​\(c\)​d~i​\(c\)​n~i​\(c\)\+ϵi,t​\(c\),\\displaystyle\+\\eta\_\{t\}\(c\)\\widetilde\{d\}\_\{i\}\(c\)\\widetilde\{n\}\_\{i\}\(c\)\+\\epsilon\_\{i,t\}\(c\),whered~i​\(c\)\\widetilde\{d\}\_\{i\}\(c\)andn~i​\(c\)\\widetilde\{n\}\_\{i\}\(c\)are the standardized log data size and log model size among configurations available at slicecc\. The main coefficients are averaged uniformly over the common compute grid:

β¯D,t=1\|𝒞\|​∑c∈𝒞βD,t​\(c\),β¯N,t=1\|𝒞\|​∑c∈𝒞βN,t​\(c\)\.\\bar\{\\beta\}\_\{D,t\}=\\frac\{1\}\{\|\\mathcal\{C\}\|\}\\sum\_\{c\\in\\mathcal\{C\}\}\\beta\_\{D,t\}\(c\),\\qquad\\bar\{\\beta\}\_\{N,t\}=\\frac\{1\}\{\|\\mathcal\{C\}\|\}\\sum\_\{c\\in\\mathcal\{C\}\}\\beta\_\{N,t\}\(c\)\.Their contrast isΔt=β¯N,t−β¯D,t\\Delta\_\{t\}=\\bar\{\\beta\}\_\{N,t\}\-\\bar\{\\beta\}\_\{D,t\}\. A negative value means that the fitted data\-size relationship is larger; it does not by itself imply that either relationship is positive\.

We use 20,000 Bayesian\-bootstrap replicates to test whether this contrast depends on the represented data sizes, model sizes, and compute slices\. Each replicate reweights those three observed axes, refits the supported slices, and recomputesΔt\\Delta\_\{t\}\. The resulting interval measures sensitivity to the observed resource conditions\. It does not include independent pretraining\-run or probe\-seed uncertainty and is not adjusted for comparisons across tasks\.

Table 11:Task\-wise relationships at matched compute\.The contrast is model size minus data size\. Intervals below zero favor pretraining data; intervals crossing zero are unresolved\.As a sensitivity check, raising the lower end of the averaging range from 0\.11 to 0\.20 EFLOP leaves ABIDE\-Age, ABIDE\-Dx, AD\-Dx, NKI\-Age, NKI\-Edu\., and SALD\-Age data\-size\-favored\. The other six intervals cross zero, and none becomes model\-size\-favored\.

### A\.4ID\-Guided Selection under Fixed Compute

The response is fitted to the equal\-weight mean validation performance across five ID tasks: ABIDE age, ABIDE diagnosis, HCP sex, PNC sex, and PPMI diagnosis\. The OOD tasks are included in the descriptive scaling analyses in Section[5](https://arxiv.org/html/2609.27232#S5), but none of their scores enters the response fit, compute budgets, or model selection\. The response follows the effective\-data family of[Muennighoff et al\. \(2023\)](https://arxiv.org/html/2609.27232#bib.bib12), with coefficients re\-estimated for fMRI downstream error and compute taken from Appendix[A\.1](https://arxiv.org/html/2609.27232#A1.SS1)\.

#### Response, Fitting, and Formula Comparison

For each saved model, we average the five ID scores and use one minus this average as ID error\. A trailing five\-model median is applied within each training run\. Because it uses only the current and preceding saved models, it does not leak later performance into an earlier compute budget\. Runs receive equal total weight so that densely saved trajectories do not dominate\.

LetDDbe pretraining data size,NNmodel size, andeeepoch\-equivalent exposure\. The number of repeated passes beyond the first isRD=e−1R\_\{D\}=e\-1\. Following Eq\. 14 of[Muennighoff et al\. \(2023\)](https://arxiv.org/html/2609.27232#bib.bib12), repeated exposure is represented by

D′=D\[1\+RD⋆\(1−exp\(−RD/RD⋆\)\)\],D^\{\\prime\}=D\\left\[1\+R\_\{D\}^\{\\star\}\\left\(1\-\\exp\(\-R\_\{D\}/R\_\{D\}^\{\\star\}\)\\right\)\\right\],where the positive parameterRD⋆R\_\{D\}^\{\\star\}controls how quickly the added value of repeated passes decreases\. The model size supported by the fitted single\-pass relation is defined using positive scale coefficientsANA\_\{N\}andBDB\_\{D\}and positive exponentsαN\\alpha\_\{N\}andβD\\beta\_\{D\}:

Nopt​\(D\)=\[αN​ANβD​BD​DβD\]1/αN\.N\_\{\\mathrm\{opt\}\}\(D\)=\\left\[\\frac\{\\alpha\_\{N\}A\_\{N\}\}\{\\beta\_\{D\}B\_\{D\}\}D^\{\\beta\_\{D\}\}\\right\]^\{1/\\alpha\_\{N\}\}\.Defining

UN=min⁡\{Nopt​\(D\),N\},RN=max⁡\{N/UN−1,0\},U\_\{N\}=\\min\\\{N\_\{\\mathrm\{opt\}\}\(D\),N\\\},\\quad R\_\{N\}=\\max\\\{N/U\_\{N\}\-1,0\\\},the effective model size is

N′=UN\[1\+RN⋆\(1−exp\(−RN/RN⋆\)\)\]\.N^\{\\prime\}=U\_\{N\}\\left\[1\+R\_\{N\}^\{\\star\}\\left\(1\-\\exp\(\-R\_\{N\}/R\_\{N\}^\{\\star\}\)\\right\)\\right\]\.The fitted ID response is

E^ID​\(D,N,e\)=E∞\+AN\(N′\)αN\+BD\(D′\)βD,\\widehat\{E\}\_\{\\mathrm\{ID\}\}\(D,N,e\)=E\_\{\\infty\}\+\\frac\{A\_\{N\}\}\{\(N^\{\\prime\}\)^\{\\alpha\_\{N\}\}\}\+\\frac\{B\_\{D\}\}\{\(D^\{\\prime\}\)^\{\\beta\_\{D\}\}\},whereE∞E\_\{\\infty\}is the empirical error floor andRN⋆R\_\{N\}^\{\\star\}is the positive capacity\-decay parameter in the definition ofN′N^\{\\prime\}\. This response separates pretraining data, model size, and repeated exposure, but is used only to compare models inside the observed range\. LetCiC\_\{i\}be the estimated pretraining compute of modelii\. At budgetC0C\_\{0\}, the selection rule is

i⋆\(C0\)=arg⁡mini:Ci≤C0E^ID\(Di,Ni,ei\)\.i^\{\\star\}\(C\_\{0\}\)=\\underset\{i:\\,C\_\{i\}\\leq C\_\{0\}\}\{\\arg\\min\}\\;\\widehat\{E\}\_\{\\mathrm\{ID\}\}\(D\_\{i\},N\_\{i\},e\_\{i\}\)\.
The fit contains 42 complete training runs, 4,125 saved models, six data sizes, and seven model sizes\. We divideDDandNNby their geometric means \(10,750\.24 recordings and 16\.561 million parameters\) for numerical conditioning\. All seven positive parameters are optimized in log space with a run\-balanced Huber objective and multistart L\-BFGS\-B\. The fMRI grid does not contain a complete single\-pass experiment: the earliest evaluated state of a run occurs after 3–154 epoch\-equivalent exposures\. The result is therefore a joint within\-grid fit of the response family, not a replication of the original single\-pass fitting protocol\.

Table 12:Response comparison and grouped holdouts\.Complete training runs, data sizes, or model sizes are removed as units\. Tail prediction fits the early portion of each run and evaluates its later models\. Regret is the ID\-score gap from the best observed model at matched compute\.
On complete\-run holdout, Eq\. 14 reduces RMSE by 33\.4% relative to the model that treats every repeated pass as new data and improves all six grouped folds \(p=0\.0156p=0\.0156, one\-sided paired Wilcoxon test\)\. Most of the improvement comes from modeling the diminishing value of repeated data passes\. Adding the capacity\-decay term to the data\-only response changes RMSE by only 2\.9%\. Run\-cluster bootstrap estimates placeRN⋆R\_\{N\}^\{\\star\}at its upper search bound in 73% of samples, so the present model range does not identify a finite capacity\-decay scale\. Raw versus smoothed trajectories and run\-balanced versus unweighted objectives give similar RMSE and data\-decay estimates\. These diagnostics support the repeated\-data correction but do not justify interpreting either decay scale as a universal number of useful passes or capacity multiples\.

#### Selected Models

After grouped validation, the response is refit on all eligible ID results\. Under the strict constraintsC≤0\.128C\\leq 0\.128andC≤1\.97C\\leq 1\.97EFLOP, it selects the models in Table\. Both budgets use the same selection rule; ties are resolved in favor of lower compute\. The fitted response and observed smoothed ID performance identify the same model at each budget\.

#### Two\-Budget OOD Audit

A training run is eligible at a budget if it contains at least one saved model whose estimated compute does not exceed that budget\. Within each eligible run, we retain the saved model with the highest predicted ID score; exact ties favor lower compute\. The audit score is the mean OOD test score across five evaluation runs, with equal weight given to AD diagnosis, ADHD diagnosis, BHRC sex, ADNI\-MCI diagnosis, NKI age, and NKI education\. Classification uses macro\-F1 and regression uses Pearson correlation\. Regret is the difference from the best eligible audit score at the same budget\.

This archived six\-task audit is a candidate\-level diagnostic and differs from the external benchmark aggregate in Figure[5](https://arxiv.org/html/2609.27232#S6.F5), which includes SALD age instead of ADNI\-MCI and uses classification accuracy\. Neither audit enters the response fit or the selection of the two final models\.

Table 14:OOD audit of the two ID\-selected models\.Rank and regret are computed among the eligible run\-level candidates at the same compute budget\. Higher score, lower rank, and lower regret are better\.
The compact audit above summarizes the available eligible\-candidate comparison\. Separate numerical outputs for the data\-heavy, model\-heavy, longer\-training, and best\-observed\-ID heuristic rules were not retained, so we do not report unverified rule\-specific values here\.

#### Additional ADNI\-MCI Results

ADNI\-MCI uses the same frozen\-probe protocol as the other OOD classification tasks\. Its full baseline comparison is reported here, rather than in the main evaluation table, solely because of the main\-paper page limit\.

ModelParametersPretrain EFLOPsAccuracyMacro\-F1BrainLM\-13M13M0\.83550\.7±6\.050\.7\\pm 6\.050\.7±6\.050\.7\\pm 6\.0BrainLM\-111M111M206\.7757\.6±5\.257\.6\\pm 5\.257\.5±5\.157\.5\\pm 5\.1BrainMASS14\.4M2\.75553\.8±3\.053\.8\\pm 3\.053\.4±2\.953\.4\\pm 2\.9BrainGFM10\.3M0\.12951\.7±7\.751\.7\\pm 7\.749\.3±6\.449\.3\\pm 6\.4BrainHarmonix\-F85\.25M110\.9057\.2±4\.357\.2\\pm 4\.357\.2±4\.257\.2\\pm 4\.2Brain\-Semantoks129\.21M0\.175559\.7±2\.159\.7\\pm 2\.159\.4±2\.259\.4\\pm 2\.2NeuroJEPA2\.65M0\.12857\.2±3\.057\.2\\pm 3\.057\.2±2\.957\.2\\pm 2\.9NeuroJEPA10\.42M1\.9754\.8±6\.754\.8\\pm 6\.754\.7±7\.054\.7\\pm 7\.0
Table 15:ADNI\-MCI frozen\-probe results\.Classification metrics are percentages\. Entries are mean±\\pmstandard deviation across five evaluation seeds; baseline results use official released weights\.

#### Baseline Compute Reconstruction

For each released baseline, we instantiate the official architecture with its pretraining input shape and measure the training FLOPs of one example\. LetFexF\_\{\\mathrm\{ex\}\}be this cost, including forward and backward computation,BBthe global batch size, andKKthe number of optimizer steps\. When the release reportsKK, compute is

Cbase=Fex​B​K/1018\.C\_\{\\mathrm\{base\}\}=F\_\{\\mathrm\{ex\}\}BK/10^\{18\}\.When it instead reportsnnexamples andEEepochs, we reconstructK=E​⌊n/B⌋K=E\\lfloor n/B\\rfloor\. An official cumulative FLOP counter takes precedence over either reconstruction\. Checkpoint metadata take precedence over configuration files, released code, and paper descriptions; parameter count alone is not used as a compute estimate\.

Table 16:Reconstructed baseline pretraining compute\.BrainLM\-111M uses its official cumulative counter\.
Schedules and metadata follow the official releases\. These reconstructions do not imply identical hardware utilization or implementation efficiency\. The strict model\-selection budgets permit no compute overrun\. Available training descriptions determine the precision of each baseline estimate, and unresolved pretraining\-cohort overlap is treated as a limitation rather than evidence of compute equivalence\.

## Appendix BBaseline Models

The public baselines span ROI sequences, functional connectomes, atlas\-derived graphs, and multimodal inputs\. We use each model’s official weights and required preprocessing\.

##### BrainLM\.

BrainLM\([186](https://arxiv.org/html/2609.27232#biba.bib186)\)learns from ROI\-level time series by predicting masked regional activity from visible context\.

##### BrainMASS\.

BrainMASS\([202](https://arxiv.org/html/2609.27232#biba.bib202)\)represents BOLD recordings as functional connectomes and aligns augmented brain networks during masked\-ROI pretraining\.

##### BrainGFM\.

BrainGFM\([199](https://arxiv.org/html/2609.27232#biba.bib199)\)learns from atlas\-derived brain graphs using contrastive learning, masked graph autoencoding, and learned prompts\.

##### BrainHarmonix\.

BrainHarmonix\([11](https://arxiv.org/html/2609.27232#biba.bib11)\)separately pretrains structural and functional MRI before fusing them through shared brain\-hub tokens\.

##### Brain\-Semantoks\.

Brain\-Semantoks\([14](https://arxiv.org/html/2609.27232#biba.bib14)\)combines semantic tokenization and self\-distillation to model regional fMRI dynamics\.

## Appendix CDataset Catalogue

Table[17](https://arxiv.org/html/2609.27232#A3.T17)lists the 201 curated sources\. We retain collections with resting\-state fMRI and remove task\-only datasets\. Counts follow dataset\-level curation, and one participant may contribute multiple recordings\. Downstream validation and test participants, together with the OOD datasets in Section[6](https://arxiv.org/html/2609.27232#S6), are excluded from the two final models’ pretraining data\. The rows remain here for documentation and link to the primary publication or repository record\.

Table 17:Curated fMRI sources\.Counts precede the nested pretraining subsets and downstream exclusions\.IDDataset and referenceSubjectsRecordings1ABCD\([7](https://arxiv.org/html/2609.27232#biba.bib7)\)810181012ABIDE\([9](https://arxiv.org/html/2609.27232#biba.bib9)\)8718713ADHD\-200\([196](https://arxiv.org/html/2609.27232#biba.bib196)\)69610614ADNI\([19](https://arxiv.org/html/2609.27232#biba.bib19)\)4974975AOMIC \(PIOP1\)\([192](https://arxiv.org/html/2609.27232#biba.bib192)\)2062066AOMIC \(PIOP2\)\([192](https://arxiv.org/html/2609.27232#biba.bib192)\)2242247BHRC\([190](https://arxiv.org/html/2609.27232#biba.bib190)\)4654658Caltech Conte Center\([21](https://arxiv.org/html/2609.27232#biba.bib21)\)1023569CCNP\([188](https://arxiv.org/html/2609.27232#biba.bib188)\)19338110CHCP\([13](https://arxiv.org/html/2609.27232#biba.bib13)\)30430411CoRR\([204](https://arxiv.org/html/2609.27232#biba.bib204)\)434101612Emo\-FilM\([25](https://arxiv.org/html/2609.27232#biba.bib25)\)96513FCON\([5](https://arxiv.org/html/2609.27232#biba.bib5)\)484814HBN\([2](https://arxiv.org/html/2609.27232#biba.bib2)\)1342134215HCP\([197](https://arxiv.org/html/2609.27232#biba.bib197)\)1011404416ISYB\([12](https://arxiv.org/html/2609.27232#biba.bib12)\)18718717MDD\([201](https://arxiv.org/html/2609.27232#biba.bib201)\)3525352518NKI\([27](https://arxiv.org/html/2609.27232#biba.bib27)\)71771719PNC\([191](https://arxiv.org/html/2609.27232#biba.bib191)\)1268126820PPMI\([23](https://arxiv.org/html/2609.27232#biba.bib23)\)47447421SALD\([198](https://arxiv.org/html/2609.27232#biba.bib198)\)49349322SLIM\([22](https://arxiv.org/html/2609.27232#biba.bib22)\)133256237T Resting\-state\([203](https://arxiv.org/html/2609.27232#biba.bib203)\)8926724ABIDE II\([10](https://arxiv.org/html/2609.27232#biba.bib10)\)1050141225AHDC\([159](https://arxiv.org/html/2609.27232#biba.bib159)\)262245126Aging MultiEcho\([193](https://arxiv.org/html/2609.27232#biba.bib193)\)298178827ALS\-FTD UHF\([176](https://arxiv.org/html/2609.27232#biba.bib176)\)333328AOMIC \(ID1000\)\([192](https://arxiv.org/html/2609.27232#biba.bib192)\)88188129Anxiety CBT Rest\([161](https://arxiv.org/html/2609.27232#biba.bib161)\)8342330Aphasia Recovery Cohort\([127](https://arxiv.org/html/2609.27232#biba.bib127)\)19250831Bilingual Rest\([50](https://arxiv.org/html/2609.27232#biba.bib50)\)929232Bilingualism Brain\([52](https://arxiv.org/html/2609.27232#biba.bib52)\)646433BOLD Variability During Cognitive Control\([140](https://arxiv.org/html/2609.27232#biba.bib140)\)15815834BTC Preop\([45](https://arxiv.org/html/2609.27232#biba.bib45)\)363635C\-PRO\([18](https://arxiv.org/html/2609.27232#biba.bib18)\)9686436CO2 Resting\([144](https://arxiv.org/html/2609.27232#biba.bib144)\)353537Cognitive Control Theoretic Mechanisms\([100](https://arxiv.org/html/2609.27232#biba.bib100)\)737338Cognitive Training\([20](https://arxiv.org/html/2609.27232#biba.bib20)\)426939Complex Multi\-Echo\([24](https://arxiv.org/html/2609.27232#biba.bib24)\)83298840Dystonia Rest\([166](https://arxiv.org/html/2609.27232#biba.bib166)\)909041Emotional Learning Rest\([109](https://arxiv.org/html/2609.27232#biba.bib109)\)308742Experience Sampling\([110](https://arxiv.org/html/2609.27232#biba.bib110)\)363643Food Brain Study\([121](https://arxiv.org/html/2609.27232#biba.bib121)\)828244GRACE\([184](https://arxiv.org/html/2609.27232#biba.bib184)\)13713745Gut\-Brain Axis\([120](https://arxiv.org/html/2609.27232#biba.bib120)\)888846Head Motion Intervention\([43](https://arxiv.org/html/2609.27232#biba.bib43)\)2416847Hearing Loss Connectome\([132](https://arxiv.org/html/2609.27232#biba.bib132)\)828248Hemodynamic Timing\([119](https://arxiv.org/html/2609.27232#biba.bib119)\)153049Hippocampal Memory Rest\([177](https://arxiv.org/html/2609.27232#biba.bib177)\)525250HRV Biofeedback\([85](https://arxiv.org/html/2609.27232#biba.bib85)\)326451HRV Biofeedback \(ds003823\)\([99](https://arxiv.org/html/2609.27232#biba.bib99)\)17697852Human Es\-fMRI\([73](https://arxiv.org/html/2609.27232#biba.bib73)\)208053Inhibitory Control Youth\([153](https://arxiv.org/html/2609.27232#biba.bib153)\)12437254Insulin Menstrual Rest\([173](https://arxiv.org/html/2609.27232#biba.bib173)\)156055IQSEC2\([91](https://arxiv.org/html/2609.27232#biba.bib91)\)2663456LEMON\([4](https://arxiv.org/html/2609.27232#biba.bib4)\)22622657Large\-scale Pre/Post\-surgical Patients\([151](https://arxiv.org/html/2609.27232#biba.bib151)\)23423458MBSR\([130](https://arxiv.org/html/2609.27232#biba.bib130)\)14734759MND\([156](https://arxiv.org/html/2609.27232#biba.bib156)\)5923660Monash RsPETMR\([76](https://arxiv.org/html/2609.27232#biba.bib76)\)2716261MSC\([15](https://arxiv.org/html/2609.27232#biba.bib15)\)1010062Multi\-echo Cambridge\([44](https://arxiv.org/html/2609.27232#biba.bib44)\)8935563Multiband Acceleration\([94](https://arxiv.org/html/2609.27232#biba.bib94)\)3231664MyConnectome\([187](https://arxiv.org/html/2609.27232#biba.bib187)\)113265Neurocon\([6](https://arxiv.org/html/2609.27232#biba.bib6)\)438566NeuroEmo\([1](https://arxiv.org/html/2609.27232#biba.bib1)\)408067NIMH Healthy Research Volunteer Dataset\([152](https://arxiv.org/html/2609.27232#biba.bib152)\)21998468Olfactory Meningioma\([181](https://arxiv.org/html/2609.27232#biba.bib181)\)565669Oxytocin\([123](https://arxiv.org/html/2609.27232#biba.bib123)\)878770Pediatric Anxiety Rest\([165](https://arxiv.org/html/2609.27232#biba.bib165)\)14954671Penn LEAD\([178](https://arxiv.org/html/2609.27232#biba.bib178)\)12741472Precision Aging Network\([183](https://arxiv.org/html/2609.27232#biba.bib183)\)69769773Prototype/Exemplar Aging\([170](https://arxiv.org/html/2609.27232#biba.bib170)\)646474QTAB\([194](https://arxiv.org/html/2609.27232#biba.bib194)\)413201475QTIM\([29](https://arxiv.org/html/2609.27232#biba.bib29)\)1195245376RT Disengagement\([169](https://arxiv.org/html/2609.27232#biba.bib169)\)595977Sleep EEG\-fMRI\([16](https://arxiv.org/html/2609.27232#biba.bib16)\)3325578Sleepy Brain\([35](https://arxiv.org/html/2609.27232#biba.bib35)\)8516579SONG\([30](https://arxiv.org/html/2609.27232#biba.bib30)\)275480SRPBS Multi\-disorder MRI\([195](https://arxiv.org/html/2609.27232#biba.bib195)\)1020102081SUDMEX\-CONN\([3](https://arxiv.org/html/2609.27232#biba.bib3)\)14114182Synaesthesia\([189](https://arxiv.org/html/2609.27232#biba.bib189)\)12750883T1 Diabetes EF Rest\([32](https://arxiv.org/html/2609.27232#biba.bib32)\)646484Tao Wu Parkinson’s Dataset\([200](https://arxiv.org/html/2609.27232#biba.bib200)\)404085Temporal Lobe Epilepsy \- UNAM\([114](https://arxiv.org/html/2609.27232#biba.bib114)\)656586THINGS\-fMRI\([17](https://arxiv.org/html/2609.27232#biba.bib17)\)335487Tinnitus\([75](https://arxiv.org/html/2609.27232#biba.bib75)\)383888TMS\-fMRI\([31](https://arxiv.org/html/2609.27232#biba.bib31)\)14815189Transdiagnostic Connectome Project\([8](https://arxiv.org/html/2609.27232#biba.bib8)\)241161390Tumor Patients Task/Rest\([129](https://arxiv.org/html/2609.27232#biba.bib129)\)4215691VASO Pulsatility Rest\([33](https://arxiv.org/html/2609.27232#biba.bib33)\)2314292Wakayama PsyRS\([26](https://arxiv.org/html/2609.27232#biba.bib26)\)20058393WashU120\([41](https://arxiv.org/html/2609.27232#biba.bib41)\)12020394Yale NeuroConnect\([179](https://arxiv.org/html/2609.27232#biba.bib179)\)41081595Yale Pupillometry\([28](https://arxiv.org/html/2609.27232#biba.bib28)\)275496Yale Reading\([141](https://arxiv.org/html/2609.27232#biba.bib141)\)8914397DMT\-HAR\-MED\([168](https://arxiv.org/html/2609.27232#biba.bib168)\)4080985\.0T Visual Scene\([182](https://arxiv.org/html/2609.27232#biba.bib182)\)2068099Frontoparietal Plasticity\([102](https://arxiv.org/html/2609.27232#biba.bib102)\)92184100Habit Learning\([113](https://arxiv.org/html/2609.27232#biba.bib113)\)1231356101NIMH CAT\-D\([117](https://arxiv.org/html/2609.27232#biba.bib117)\)1302369102NIMH Ketamine Mechanism of Action Study\([160](https://arxiv.org/html/2609.27232#biba.bib160)\)581471103NIMH METeR\([126](https://arxiv.org/html/2609.27232#biba.bib126)\)5148104NODEAP\([171](https://arxiv.org/html/2609.27232#biba.bib171)\)48996105Infra\-Low Frequency Neurofeedback\([108](https://arxiv.org/html/2609.27232#biba.bib108)\)918106ON\-Harmony\([122](https://arxiv.org/html/2609.27232#biba.bib122)\)20184107PAFIN\([163](https://arxiv.org/html/2609.27232#biba.bib163)\)511020108PLP NF1\([80](https://arxiv.org/html/2609.27232#biba.bib80)\)18715109PSYCH\-REST\([175](https://arxiv.org/html/2609.27232#biba.bib175)\)34271110Pragmatic Language\([93](https://arxiv.org/html/2609.27232#biba.bib93)\)145572111PsiConnect\([162](https://arxiv.org/html/2609.27232#biba.bib162)\)652020112RPN Signature Study 1\([64](https://arxiv.org/html/2609.27232#biba.bib64)\)4141113RPN Signature Study 2\([65](https://arxiv.org/html/2609.27232#biba.bib65)\)4949114Reading Brain Project L1 Adults\([104](https://arxiv.org/html/2609.27232#biba.bib104)\)52312115Reading Brain Project L2 Adults\([105](https://arxiv.org/html/2609.27232#biba.bib105)\)56336116Reinforcement\-Learning Generalization\([138](https://arxiv.org/html/2609.27232#biba.bib138)\)44130117SUDMEX\-TMS\([81](https://arxiv.org/html/2609.27232#biba.bib81)\)53154118Speech Disfluencies\([92](https://arxiv.org/html/2609.27232#biba.bib92)\)8181119SpiDa\-MRI\([118](https://arxiv.org/html/2609.27232#biba.bib118)\)49343120Temporal Dynamics of Emotional Music\([82](https://arxiv.org/html/2609.27232#biba.bib82)\)39156121Truecrime\([128](https://arxiv.org/html/2609.27232#biba.bib128)\)133133122Udall Pilot ANT\([53](https://arxiv.org/html/2609.27232#biba.bib53)\)9126123Valenced Tactile Information\([145](https://arxiv.org/html/2609.27232#biba.bib145)\)1151830124Visuomotor Rotation Adaptation Experiment\([107](https://arxiv.org/html/2609.27232#biba.bib107)\)32320125Visuomotor Rotation Learning and Reward\-based Motor Learning\([149](https://arxiv.org/html/2609.27232#biba.bib149)\)45308126Resting State and Arithmetic Task\([63](https://arxiv.org/html/2609.27232#biba.bib63)\)46138127Fibromyalgia Emotion Regulation Dataset\([111](https://arxiv.org/html/2609.27232#biba.bib111)\)66131128Closed\-eyes Depression and Healthy Controls\([70](https://arxiv.org/html/2609.27232#biba.bib70)\)7272129HC / PD\-NC / PD\-MCI Resting\-State MRI\([158](https://arxiv.org/html/2609.27232#biba.bib158)\)5555130rewardBeast\([69](https://arxiv.org/html/2609.27232#biba.bib69)\)35121131MULTI\-CLARID\([154](https://arxiv.org/html/2609.27232#biba.bib154)\)3439132Modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fMRI study\([34](https://arxiv.org/html/2609.27232#biba.bib34)\)26156133Brain connectivity predicts placebo response across chronic pain clinical trials\([36](https://arxiv.org/html/2609.27232#biba.bib36)\)7676134Multi\-echo fMRI replication sample of autobiographical memory, prospection and theory of mind reasoning tasks\([37](https://arxiv.org/html/2609.27232#biba.bib37)\)3193135Multiband Multi\-Echo Imaging of Simultaneous Oxygenation and Flow Timeseries for Resting State Connectivity\([38](https://arxiv.org/html/2609.27232#biba.bib38)\)728136Cost Analysis TBI\([39](https://arxiv.org/html/2609.27232#biba.bib39)\)2666137MPI\-Leipzig\_Mind\-Brain\-Body\([40](https://arxiv.org/html/2609.27232#biba.bib40)\)316994138ds000245\_R1\.0\.0\([42](https://arxiv.org/html/2609.27232#biba.bib42)\)4545139rsfMRI\_single\_session\_EEG\_NF\([46](https://arxiv.org/html/2609.27232#biba.bib46)\)52104140Human hippocampal replay during rest prioritizes weakly\-learned information and predicts memory performance\([47](https://arxiv.org/html/2609.27232#biba.bib47)\)2448141Layer VASO in visual system\([48](https://arxiv.org/html/2609.27232#biba.bib48)\)48142100 runs at 3T\([49](https://arxiv.org/html/2609.27232#biba.bib49)\)319143InterTVA\. A multimodal MRI dataset for the study of inter\-individual differences in voice perception and identification\.\([51](https://arxiv.org/html/2609.27232#biba.bib51)\)4040144The physiological effects of non\-invasive brain stimulation fundamentally differ across the human cortex\([54](https://arxiv.org/html/2609.27232#biba.bib54)\)23133145Functional Connectivity of Music\-Induced Analgesia in Fibromyalgia\([55](https://arxiv.org/html/2609.27232#biba.bib55)\)40160146Auditory localization with 7T fMRI\([56](https://arxiv.org/html/2609.27232#biba.bib56)\)99147Multi\-domain task battery \(MDTB\)\([57](https://arxiv.org/html/2609.27232#biba.bib57)\)1837148Simultaneous eeg\-fmri for a speeded discrimination task with confidence\([58](https://arxiv.org/html/2609.27232#biba.bib58)\)2020149Caltech rsfMRI Dataset\([59](https://arxiv.org/html/2609.27232#biba.bib59)\)612150Dense Investigation of Variability of Affect \(DIVA\)\([60](https://arxiv.org/html/2609.27232#biba.bib60)\)3152151Neuroimaging predictors of creativity in healthy adults\([61](https://arxiv.org/html/2609.27232#biba.bib61)\)66132152Yale\_Single\_Subject\_Task\_Rest30x\([62](https://arxiv.org/html/2609.27232#biba.bib62)\)153153Meditacion Interocepcion\([66](https://arxiv.org/html/2609.27232#biba.bib66)\)1115428andMe\([67](https://arxiv.org/html/2609.27232#biba.bib67)\)159155Headmold\([68](https://arxiv.org/html/2609.27232#biba.bib68)\)1144156YanDataBIDS\([71](https://arxiv.org/html/2609.27232#biba.bib71)\)33157Cast\-induced plasticity\([72](https://arxiv.org/html/2609.27232#biba.bib72)\)3197158Aging\([74](https://arxiv.org/html/2609.27232#biba.bib74)\)3939159SoccerCAN\([77](https://arxiv.org/html/2609.27232#biba.bib77)\)2446160Two sessions of resting state with closed eyes for patients with depression in treatment course \(NFB, CBT or No treatment groups\)\([78](https://arxiv.org/html/2609.27232#biba.bib78)\)2958161Social Processes Initiative in Neurobiology of the Schizophrenia\(s\) Traveling Human Phantoms\([79](https://arxiv.org/html/2609.27232#biba.bib79)\)430162How ovarian hormones influence the behaviroal activation and inhibition system through the dopamine pathway\([83](https://arxiv.org/html/2609.27232#biba.bib83)\)4949163Using anesthesia\-induced loss of consciousness to identify biomarkers of conscious awareness in the healthy human brain\([84](https://arxiv.org/html/2609.27232#biba.bib84)\)1768164Monash vis\-fPET\-fMRI\([86](https://arxiv.org/html/2609.27232#biba.bib86)\)1030165Brain Network Mechanisms of Visual Shape Completion\([87](https://arxiv.org/html/2609.27232#biba.bib87)\)1919166Identification of an Amygdala\-Thalamic Circuit That Acts as a Central Gain Mechanism in Taste Perception\([88](https://arxiv.org/html/2609.27232#biba.bib88)\)2828167Protecting the Aging Brain \- Diet\-Study\([89](https://arxiv.org/html/2609.27232#biba.bib89)\)1236168Protecting the Aging Brain, Case\-Study\([90](https://arxiv.org/html/2609.27232#biba.bib90)\)12169Emotion Category and Face Perception Task Optimized for Multivariate Pattern Analysis\([95](https://arxiv.org/html/2609.27232#biba.bib95)\)1616170Reward biases spontaneous neural reactivation during sleep\([96](https://arxiv.org/html/2609.27232#biba.bib96)\)1818171ScanTrain\([97](https://arxiv.org/html/2609.27232#biba.bib97)\)3130172Parallel systems for social and spatial reasoning\([98](https://arxiv.org/html/2609.27232#biba.bib98)\)10300173PE\-Update\([101](https://arxiv.org/html/2609.27232#biba.bib101)\)2446174Resting\-state for 34 younger and 28 older adults\([103](https://arxiv.org/html/2609.27232#biba.bib103)\)6262175Pre\-Post rehabilitation fMRI data of post\-stroke patients\.\([106](https://arxiv.org/html/2609.27232#biba.bib106)\)2958176rest\_eye\([112](https://arxiv.org/html/2609.27232#biba.bib112)\)2040177Perinatal Stroke\([115](https://arxiv.org/html/2609.27232#biba.bib115)\)164178Ironia VEV\([116](https://arxiv.org/html/2609.27232#biba.bib116)\)4141179CS\-DSI\([124](https://arxiv.org/html/2609.27232#biba.bib124)\)2020180language fMRI\([125](https://arxiv.org/html/2609.27232#biba.bib125)\)7171181POP\([131](https://arxiv.org/html/2609.27232#biba.bib131)\)3940182The Neural Basis of Visual Shape Completion in Schizophrenia and Bipolar Disorder\([133](https://arxiv.org/html/2609.27232#biba.bib133)\)3030183Neural Correlates of Lidocaine Analgesic \(NLA\) Study\([134](https://arxiv.org/html/2609.27232#biba.bib134)\)275418428andHe\([135](https://arxiv.org/html/2609.27232#biba.bib135)\)140185Weill Cornell Medicine Multi\-echo \(WCM\-ME\) Dataset\([136](https://arxiv.org/html/2609.27232#biba.bib136)\)1220186AMRI 16\-N\-0031 sleep1\([137](https://arxiv.org/html/2609.27232#biba.bib137)\)1321187SoCal Kinesia and Incentivization for Parkinson’s Disease \(SKIP\): Ultra\-High Field Functional Connectivity\([139](https://arxiv.org/html/2609.27232#biba.bib139)\)2884188BABA: Naturalistic fMRI and MEG Dataset\([142](https://arxiv.org/html/2609.27232#biba.bib142)\)2929189Protecting the Aging Brain \- fMRI study of the brain in ketosis\([143](https://arxiv.org/html/2609.27232#biba.bib143)\)101404190Priority\([146](https://arxiv.org/html/2609.27232#biba.bib146)\)2987191Circadian misalignment and energy balance\([147](https://arxiv.org/html/2609.27232#biba.bib147)\)11176192A comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients\([148](https://arxiv.org/html/2609.27232#biba.bib148)\)1632193Mitchell\_Hacker\_2013\([150](https://arxiv.org/html/2609.27232#biba.bib150)\)853194The DBS\-fMRI dataset\([155](https://arxiv.org/html/2609.27232#biba.bib155)\)14880195China’s Social Fake News database release with brain structural, functional, and behavioural measures\([157](https://arxiv.org/html/2609.27232#biba.bib157)\)4343196QNL NegativeBOLD Database\([164](https://arxiv.org/html/2609.27232#biba.bib164)\)255290197Chicago Attention and Thoughts\([167](https://arxiv.org/html/2609.27232#biba.bib167)\)60231198Night Owls Scan Club\([172](https://arxiv.org/html/2609.27232#biba.bib172)\)4184199Linking Subjective Experience of Anxiety to Brain Function using Natural Language Processing\.\([174](https://arxiv.org/html/2609.27232#biba.bib174)\)75447200Dense longitudinal single\-subject multimodal MRI dataset acquired via self\-administered scanning\([180](https://arxiv.org/html/2609.27232#biba.bib180)\)1458201Multi\-scale, multi\-modal imaging assessment of trajectories of cognitive impairment in Multiple Sclerosis\([185](https://arxiv.org/html/2609.27232#biba.bib185)\)28270Table 17:Curated fMRI sources\(continued\)\.
## Appendix References

- \(1\)Abgeena, Shruti Garg, and Nishant Goyal\.Neuroemo: An fmri dataset for emotion recognition, 2025\.URL[https://openneuro\.org/datasets/ds005700](https://openneuro.org/datasets/ds005700)\.
- \(2\)Lindsay M\. Alexander, Jasmine Escalera, Lei Ai, et al\.An open resource for transdiagnostic research in pediatric mental health and learning disorders\.*Scientific Data*, 4\(1\):170181, 2017\.doi:10\.1038/sdata\.2017\.181\.
- \(3\)Diego Angeles\-Valdez, Jalil Rasgado\-Toledo, Victor Issa\-Garcia, Thania Balducci, Viviana Villicaña, Alely Valencia, Jorge Julio Gonzalez\-Olvera, Ernesto Reyes\-Zamorano, and Eduardo A\. Garza\-Villarreal\.The mexican magnetic resonance imaging dataset of patients with cocaine use disorder: SUDMEX CONN\.*Scientific Data*, 9\(1\):133, 2022\.doi:10\.1038/s41597\-022\-01251\-3\.URL[https://www\.nature\.com/articles/s41597\-022\-01251\-3](https://www.nature.com/articles/s41597-022-01251-3)\.
- \(4\)Anahit Babayan, Miray Erbey, Deniz Kumral, et al\.A mind\-brain\-body dataset of mri, eeg, cognition, emotion, and peripheral physiology in young and old adults\.*Scientific Data*, 6\(1\):180308, 2019\.doi:10\.1038/sdata\.2018\.308\.URL[https://www\.nature\.com/articles/sdata2018308](https://www.nature.com/articles/sdata2018308)\.
- \(5\)Bharat B\. Biswal, Maarten Mennes, Xi\-Nian Zuo, et al\.Toward discovery science of human brain function\.*Proceedings of the National Academy of Sciences*, 107\(10\):4734–4739, 2010\.doi:10\.1073/pnas\.0911855107\.
- \(6\)Canadian Open Neuroscience Platform\.Neurocon dataset, 2026\.URL[https://portal\.conp\.ca/dataset?id=projects%2FNeurocon](https://portal.conp.ca/dataset?id=projects%2FNeurocon)\.Accessed 2026\-05\-12\.
- \(7\)B\. J\. Casey, Tariq Cannonier, May I\. Conley, et al\.The adolescent brain cognitive development \(ABCD\) study: Imaging acquisition across 21 sites\.*Developmental Cognitive Neuroscience*, 32:43–54, 2018\.doi:10\.1016/j\.dcn\.2018\.03\.001\.
- \(8\)Sidhant Chopra, Carrisa V\. Cocuzza, Connor Lawhead, Jocelyn A\. Ricard, Loïc Labache, Lauren M\. Patrick, Poornima Kumar, Arielle Rubenstein, et al\.The transdiagnostic connectome project: an open dataset for studying brain\-behavior relationships in psychiatry\.*Scientific Data*, 12:923, 2025\.doi:10\.1038/s41597\-025\-04895\-z\.URL[https://www\.nature\.com/articles/s41597\-025\-04895\-z](https://www.nature.com/articles/s41597-025-04895-z)\.
- \(9\)Adriana Di Martino, Chao\-Gan Yan, Qingyang Li, et al\.The autism brain imaging data exchange: towards a large\-scale evaluation of the intrinsic brain architecture in autism\.*Molecular Psychiatry*, 19\(6\):659–667, 2014\.doi:10\.1038/mp\.2013\.78\.
- \(10\)Adriana Di Martino, David O’Connor, Bosi Chen, et al\.Enhancing studies of the connectome in autism using the autism brain imaging data exchange II\.*Scientific Data*, 4:170010, 2017\.doi:10\.1038/sdata\.2017\.10\.URL[https://www\.nature\.com/articles/sdata201710](https://www.nature.com/articles/sdata201710)\.
- \(11\)Zijian Dong, Ruilin Li, Joanna Chong, Niousha Dehestani, Yinghui Teng, Yi Lin, Zhizhou Li, Yichi Zhang, Yapei Xie, Leon Ooi, B\. T\. Thomas Yeo, and Juan Helen Zhou\.Brain harmony: A multimodal foundation model unifying morphology and function into 1d tokens\.In*Advances in Neural Information Processing Systems*, volume 38, 2025\.URL[https://papers\.neurips\.cc/paper\_files/paper/2025/hash/b0832cc57899cfd2d3fedeb3f330ba80\-Abstract\-Conference\.html](https://papers.neurips.cc/paper_files/paper/2025/hash/b0832cc57899cfd2d3fedeb3f330ba80-Abstract-Conference.html)\.
- \(12\)Peng Gao, Hao\-Ming Dong, Si\-Man Liu, et al\.A chinese multi\-modal neuroimaging data release for increasing diversity of human brain mapping\.*Scientific Data*, 9\(1\):297, 2022\.doi:10\.1038/s41597\-022\-01413\-3\.
- \(13\)Jianqiao Ge, Guangyu Yang, Meng Han, et al\.Increasing diversity in connectomics with the chinese human connectome project\.*Nature Neuroscience*, 26\(1\):163–172, 2023\.doi:10\.1038/s41593\-022\-01215\-1\.
- \(14\)Sam Gijsen, Marc\-Andre Schulz, and Kerstin Ritter\.Brain\-Semantoks: Learning semantic tokens of brain dynamics with a self\-distilled foundation model\.In*International Conference on Learning Representations*, 2026\.URL[https://iclr\.cc/virtual/2026/poster/10011042](https://iclr.cc/virtual/2026/poster/10011042)\.
- \(15\)Evan M\. Gordon, Timothy O\. Laumann, Babatunde Adeyemo, and Steven E\. Petersen\.Precision functional mapping of individual human brains\.*Neuron*, 95\(4\):791–807, 2017\.doi:10\.1016/j\.neuron\.2017\.07\.011\.URL[https://www\.cell\.com/neuron/fulltext/S0896\-6273\(17\)30613\-X](https://www.cell.com/neuron/fulltext/S0896-6273(17)30613-X)\.
- \(16\)Yameng Gu, Feng Han, Lucas E\. Sainburg, Margeaux M\. Schade, Orfeu M\. Buxton, Jeff H\. Duyn, and Xiao Liu\.Simultaneous eeg and functional mri data during rest and sleep from humans\.*Data in Brief*, 48:109059, 2023\.doi:10\.1016/j\.dib\.2023\.109059\.URL[https://www\.sciencedirect\.com/science/article/pii/S2352340923001774](https://www.sciencedirect.com/science/article/pii/S2352340923001774)\.
- \(17\)Martin N\. Hebart, Oliver Contier, Lina Teichmann, Adrian H\. Rockter, Charles Y\. Zheng, Andrew Kidder, Annie Corriveau, Lukas Muttenthaler, Alexander Walther, Arjen Alink, and Chris I\. Baker\.Things\-data, a multimodal collection of large\-scale datasets for investigating object representations in human brain and behavior\.*eLife*, 12:e82580, 2023\.doi:10\.7554/eLife\.82580\.URL[https://elifesciences\.org/articles/82580](https://elifesciences.org/articles/82580)\.
- \(18\)Takuya Ito, Guangyu Robert Yang, Patryk Laurent, Douglas H\. Schultz, and Michael W\. Cole\.Constructing neural network models from brain data reveals representational transformations linked to adaptive behavior\.*Nature Communications*, 13\(1\):673, 2022\.doi:10\.1038/s41467\-022\-28323\-7\.URL[https://www\.nature\.com/articles/s41467\-022\-28323\-7](https://www.nature.com/articles/s41467-022-28323-7)\.
- \(19\)Clifford R\. Jack, Matt A\. Bernstein, Nick C\. Fox, et al\.The alzheimer’s disease neuroimaging initiative \(ADNI\): MRI methods\.*Journal of Magnetic Resonance Imaging*, 27\(4\):685–691, 2008\.doi:10\.1002/jmri\.21049\.
- \(20\)Joseph W\. Kable, Molly K\. Caulfield, Michael Falcone, Maria McConnell, Luis Bernardo, T\. Parthasarathi, Nora Cooper, Rebecca L\. Ashare, Janet Audrain\-McGovern, Robert Hornik, Paul Diefenbach, Frank Lee, and Caryn Lerman\.No effect of commercial cognitive training on brain activity, choice behavior, or cognitive performance\.*The Journal of Neuroscience*, 37\(31\):7390–7402, 2017\.doi:10\.1523/JNEUROSCI\.2832\-16\.2017\.URL[https://www\.jneurosci\.org/content/37/31/7390](https://www.jneurosci.org/content/37/31/7390)\.
- \(21\)Dorit Kliemann, Ralph Adolphs, Tim Armstrong, et al\.Caltech conte center, a multimodal data resource for exploring social cognition and decision\-making\.*Scientific Data*, 9\(1\):138, 2022\.doi:10\.1038/s41597\-022\-01171\-2\.
- \(22\)Wei Liu, Dongtao Wei, Qunlin Chen, et al\.Longitudinal test\-retest neuroimaging data from healthy young adults in southwest china\.*Scientific Data*, 4\(1\):170017, 2017\.doi:10\.1038/sdata\.2017\.17\.
- \(23\)Kenneth Marek, Danna Jennings, Shirley Lasch, et al\.The parkinson progression marker initiative \(PPMI\)\.*Progress in Neurobiology*, 95\(4\):629–635, 2011\.doi:10\.1016/j\.pneurobio\.2011\.09\.005\.
- \(24\)Michal Mikl, Kateřina Ingrová, Martin Gajdoš, Marie Schejbalová, Anežka Kovářová, Radek Mareček, and Tomáš Slavíček\.Complex multi\-echo fmri dataset: New strategies in processing of multi\-echo data\.*Scientific Data*, 13:320, 2026\.doi:10\.1038/s41597\-026\-06694\-6\.URL[https://www\.nature\.com/articles/s41597\-026\-06694\-6](https://www.nature.com/articles/s41597-026-06694-6)\.
- \(25\)Elenor Morgenroth, Stefano Moia, Laura Vilaclara, Raphael Fournier, Michal Muszynski, Maria Ploumitsakou, Marina Almató\-Bellavista, Patrik Vuilleumier, and Dimitri Van De Ville\.Emo\-film: A multimodal dataset for affective neuroscience using naturalistic stimuli\.*Scientific Data*, 12\(1\):684, 2025\.doi:10\.1038/s41597\-025\-04803\-5\.URL[https://www\.nature\.com/articles/s41597\-025\-04803\-5](https://www.nature.com/articles/s41597-025-04803-5)\.
- \(26\)Neuroimaging Informatics Tools and Resources Clearinghouse\.Resting\-state fmri from healthy young subjects \(Wakayama\_PsyRS\), 2018\.URL[https://www\.nitrc\.org/projects/wakayama\_psyrs](https://www.nitrc.org/projects/wakayama_psyrs)\.
- \(27\)Kate Brody Nooner, Stanley J\. Colcombe, Russell H\. Tobe, et al\.The NKI\-rockland sample: A model for accelerating the pace of discovery science in psychiatry\.*Frontiers in Neuroscience*, 6:152, 2012\.doi:10\.3389/fnins\.2012\.00152\.
- \(28\)OpenNeuro\.Openneuro dataset ds003673 \(yale pupillometry\), 2026a\.URL[https://openneuro\.org/datasets/ds003673](https://openneuro.org/datasets/ds003673)\.Accessed 2026\-06\-03\.
- \(29\)OpenNeuro\.Openneuro dataset ds004169 \(QTIM\), 2026b\.URL[https://openneuro\.org/datasets/ds004169](https://openneuro.org/datasets/ds004169)\.Accessed 2026\-06\-03\.
- \(30\)OpenNeuro\.Openneuro dataset ds004592 \(SONG\), 2026c\.URL[https://openneuro\.org/datasets/ds004592](https://openneuro.org/datasets/ds004592)\.Accessed 2026\-06\-03\.
- \(31\)OpenNeuro\.Openneuro dataset ds005498 \(tms\-fmri\), 2026d\.URL[https://openneuro\.org/datasets/ds005498](https://openneuro.org/datasets/ds005498)\.Accessed 2026\-06\-03\.
- \(32\)OpenNeuro\.Openneuro dataset ds006156 \(t1 diabetes ef rest\), 2026e\.URL[https://openneuro\.org/datasets/ds006156](https://openneuro.org/datasets/ds006156)\.Accessed 2026\-06\-03\.
- \(33\)OpenNeuro\.Openneuro dataset ds006212 \(vaso pulsatility rest\), 2026f\.URL[https://openneuro\.org/datasets/ds006212](https://openneuro.org/datasets/ds006212)\.Accessed 2026\-06\-03\.
- \(34\)OpenNeuro Dataset ds000133\.Openneuro dataset ds000133 \(modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fmri study\), 2026\.URL[https://openneuro\.org/datasets/ds000133](https://openneuro.org/datasets/ds000133)\.Accessed 2026\-07\-23\.
- \(35\)OpenNeuro Dataset ds000201\.Openneuro dataset ds000201 \(sleepy brain\), 2026\.URL[https://openneuro\.org/datasets/ds000201](https://openneuro.org/datasets/ds000201)\.Accessed 2026\-06\-22\.
- \(36\)OpenNeuro Dataset ds000208\.Openneuro dataset ds000208 \(brain connectivity predicts placebo response across chronic pain clinical trials\), 2026\.URL[https://openneuro\.org/datasets/ds000208](https://openneuro.org/datasets/ds000208)\.Accessed 2026\-07\-23\.
- \(37\)OpenNeuro Dataset ds000210\.Openneuro dataset ds000210 \(multi\-echo fmri replication sample of autobiographical memory, prospection and theory of mind reasoning tasks\), 2026\.URL[https://openneuro\.org/datasets/ds000210](https://openneuro.org/datasets/ds000210)\.Accessed 2026\-07\-23\.
- \(38\)OpenNeuro Dataset ds000216\.Openneuro dataset ds000216 \(multiband multi\-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity\), 2026\.URL[https://openneuro\.org/datasets/ds000216](https://openneuro.org/datasets/ds000216)\.Associated publication DOI: 10\.1371/journal\.pone\.0169253; accessed 2026\-07\-23\.
- \(39\)OpenNeuro Dataset ds000220\.Openneuro dataset ds000220 \(cost analysis tbi\), 2026\.URL[https://openneuro\.org/datasets/ds000220](https://openneuro.org/datasets/ds000220)\.Accessed 2026\-07\-23\.
- \(40\)OpenNeuro Dataset ds000221\.Openneuro dataset ds000221 \(mpi\-leipzig\_mind\-brain\-body\), 2026\.URL[https://openneuro\.org/datasets/ds000221](https://openneuro.org/datasets/ds000221)\.Accessed 2026\-07\-23\.
- \(41\)OpenNeuro Dataset ds000243\.Washington university 120 \(WashU120\), 2026\.URL[https://openneuro\.org/datasets/ds000243](https://openneuro.org/datasets/ds000243)\.Accessed 2026\-05\-12\.
- \(42\)OpenNeuro Dataset ds000245\.Openneuro dataset ds000245 \(ds000245\_r1\.0\.0\), 2026\.URL[https://openneuro\.org/datasets/ds000245](https://openneuro.org/datasets/ds000245)\.Accessed 2026\-07\-23\.
- \(43\)OpenNeuro Dataset ds000256\.Openneuro dataset ds000256 \(head motion intervention\), 2026\.URL[https://openneuro\.org/datasets/ds000256](https://openneuro.org/datasets/ds000256)\.Accessed 2026\-06\-22\.
- \(44\)OpenNeuro Dataset ds000258\.Openneuro dataset ds000258 \(multi\-echo cambridge\), 2026\.URL[https://openneuro\.org/datasets/ds000258](https://openneuro.org/datasets/ds000258)\.Accessed 2026\-06\-03\.
- \(45\)OpenNeuro Dataset ds001226\.Openneuro dataset ds001226 \(btc preop\), 2026\.URL[https://openneuro\.org/datasets/ds001226](https://openneuro.org/datasets/ds001226)\.Accessed 2026\-06\-22\.
- \(46\)OpenNeuro Dataset ds001408\.Openneuro dataset ds001408 \(rsfmri\_single\_session\_eeg\_nf\), 2026\.URL[https://openneuro\.org/datasets/ds001408](https://openneuro.org/datasets/ds001408)\.Accessed 2026\-07\-23\.
- \(47\)OpenNeuro Dataset ds001454\.Openneuro dataset ds001454 \(human hippocampal replay during rest prioritizes weakly\-learned information and predicts memory performance\), 2026\.URL[https://openneuro\.org/datasets/ds001454](https://openneuro.org/datasets/ds001454)\.Accessed 2026\-07\-23\.
- \(48\)OpenNeuro Dataset ds001547\.Openneuro dataset ds001547 \(layer vaso in visual system\), 2026\.URL[https://openneuro\.org/datasets/ds001547](https://openneuro.org/datasets/ds001547)\.Accessed 2026\-07\-23\.
- \(49\)OpenNeuro Dataset ds001553\.Openneuro dataset ds001553 \(100 runs at 3t\), 2026\.URL[https://openneuro\.org/datasets/ds001553](https://openneuro.org/datasets/ds001553)\.Associated publication DOI: 10\.1073/pnas\.1121049109; accessed 2026\-07\-23\.
- \(50\)OpenNeuro Dataset ds001747\.Openneuro dataset ds001747 \(bilingual rest\), 2026\.URL[https://openneuro\.org/datasets/ds001747](https://openneuro.org/datasets/ds001747)\.Accessed 2026\-06\-22\.
- \(51\)OpenNeuro Dataset ds001771\.Openneuro dataset ds001771 \(intertva\. a multimodal mri dataset for the study of inter\-individual differences in voice perception and identification\.\), 2026\.URL[https://openneuro\.org/datasets/ds001771](https://openneuro.org/datasets/ds001771)\.Accessed 2026\-07\-23\.
- \(52\)OpenNeuro Dataset ds001796\.Openneuro dataset ds001796 \(bilingualism brain\), 2026\.URL[https://openneuro\.org/datasets/ds001796](https://openneuro.org/datasets/ds001796)\.Accessed 2026\-06\-22\.
- \(53\)OpenNeuro Dataset ds001907\.Openneuro dataset ds001907 \(udall pilot ant\), 2026\.URL[https://openneuro\.org/datasets/ds001907](https://openneuro.org/datasets/ds001907)\.Accessed 2026\-06\-22\.
- \(54\)OpenNeuro Dataset ds001927\.Openneuro dataset ds001927 \(the physiological effects of non\-invasive brain stimulation fundamentally differ across the human cortex\), 2026\.URL[https://openneuro\.org/datasets/ds001927](https://openneuro.org/datasets/ds001927)\.Associated publication DOI: 10\.1126/sciadv\.aay2739; accessed 2026\-07\-23\.
- \(55\)OpenNeuro Dataset ds001928\.Openneuro dataset ds001928 \(functional connectivity of music\-induced analgesia in fibromyalgia\), 2026\.URL[https://openneuro\.org/datasets/ds001928](https://openneuro.org/datasets/ds001928)\.Accessed 2026\-07\-23\.
- \(56\)OpenNeuro Dataset ds001942\.Openneuro dataset ds001942 \(auditory localization with 7t fmri\), 2026\.URL[https://openneuro\.org/datasets/ds001942](https://openneuro.org/datasets/ds001942)\.Accessed 2026\-07\-23\.
- \(57\)OpenNeuro Dataset ds002105\.Openneuro dataset ds002105 \(multi\-domain task battery \(mdtb\)\), 2026\.URL[https://openneuro\.org/datasets/ds002105](https://openneuro.org/datasets/ds002105)\.Associated publication DOI: 10\.1038/s41593\-019\-0436\-x; accessed 2026\-07\-23\.
- \(58\)OpenNeuro Dataset ds002158\.Openneuro dataset ds002158 \(simultaneous eeg\-fmri for a speeded discrimination task with confidence\), 2026\.URL[https://openneuro\.org/datasets/ds002158](https://openneuro.org/datasets/ds002158)\.Associated publication DOI: 10\.1073/pnas\.1918335117; accessed 2026\-07\-23\.
- \(59\)OpenNeuro Dataset ds002232\.Openneuro dataset ds002232 \(caltech rsfmri dataset\), 2026\.URL[https://openneuro\.org/datasets/ds002232](https://openneuro.org/datasets/ds002232)\.Accessed 2026\-07\-23\.
- \(60\)OpenNeuro Dataset ds002278\.Openneuro dataset ds002278 \(dense investigation of variability of affect \(diva\)\), 2026\.URL[https://openneuro\.org/datasets/ds002278](https://openneuro.org/datasets/ds002278)\.Accessed 2026\-07\-23\.
- \(61\)OpenNeuro Dataset ds002330\.Openneuro dataset ds002330 \(neuroimaging predictors of creativity in healthy adults\), 2026\.URL[https://openneuro\.org/datasets/ds002330](https://openneuro.org/datasets/ds002330)\.Associated publication DOI: 10\.1016/j\.neuroimage\.2019\.116292; accessed 2026\-07\-23\.
- \(62\)OpenNeuro Dataset ds002372\.Openneuro dataset ds002372 \(yale\_single\_subject\_task\_rest30x\), 2026\.URL[https://openneuro\.org/datasets/ds002372](https://openneuro.org/datasets/ds002372)\.Accessed 2026\-07\-23\.
- \(63\)OpenNeuro Dataset ds002422\.Openneuro dataset ds002422 \(resting state and arithmetic task\), 2026\.URL[https://openneuro\.org/datasets/ds002422](https://openneuro.org/datasets/ds002422)\.Accessed 2026\-06\-22\.
- \(64\)OpenNeuro Dataset ds002608\.Openneuro dataset ds002608 \(rpn signature study 1\), 2026\.URL[https://openneuro\.org/datasets/ds002608](https://openneuro.org/datasets/ds002608)\.Accessed 2026\-06\-22\.
- \(65\)OpenNeuro Dataset ds002609\.Openneuro dataset ds002609 \(rpn signature study 2\), 2026\.URL[https://openneuro\.org/datasets/ds002609](https://openneuro.org/datasets/ds002609)\.Accessed 2026\-06\-22\.
- \(66\)OpenNeuro Dataset ds002614\.Openneuro dataset ds002614 \(meditacion interocepcion\), 2026\.URL[https://openneuro\.org/datasets/ds002614](https://openneuro.org/datasets/ds002614)\.Accessed 2026\-07\-23\.
- \(67\)OpenNeuro Dataset ds002674\.Openneuro dataset ds002674 \(28andme\), 2026\.URL[https://openneuro\.org/datasets/ds002674](https://openneuro.org/datasets/ds002674)\.Associated publication DOI: 10\.1101/866913; accessed 2026\-07\-23\.
- \(68\)OpenNeuro Dataset ds002735\.Openneuro dataset ds002735 \(headmold\), 2026\.URL[https://openneuro\.org/datasets/ds002735](https://openneuro.org/datasets/ds002735)\.Associated publication DOI: 10\.1016/j\.neuroimage\.2019\.01\.016; accessed 2026\-07\-23\.
- \(69\)OpenNeuro Dataset ds002738\.Openneuro dataset ds002738 \(rewardbeast\), 2026\.URL[https://openneuro\.org/datasets/ds002738](https://openneuro.org/datasets/ds002738)\.Accessed 2026\-06\-22\.
- \(70\)OpenNeuro Dataset ds002748\.Openneuro dataset ds002748 \(closed\-eyes depression and healthy controls\), 2026\.URL[https://openneuro\.org/datasets/ds002748](https://openneuro.org/datasets/ds002748)\.Accessed 2026\-06\-22\.
- \(71\)OpenNeuro Dataset ds002750\.Openneuro dataset ds002750 \(yandatabids\), 2026\.URL[https://openneuro\.org/datasets/ds002750](https://openneuro.org/datasets/ds002750)\.Accessed 2026\-07\-23\.
- \(72\)OpenNeuro Dataset ds002766\.Openneuro dataset ds002766 \(cast\-induced plasticity\), 2026\.URL[https://openneuro\.org/datasets/ds002766](https://openneuro.org/datasets/ds002766)\.Associated publication DOI: 10\.1016/j\.neuron\.2020\.05\.007; accessed 2026\-07\-23\.
- \(73\)OpenNeuro Dataset ds002799\.Openneuro dataset ds002799 \(human es\-fmri\), 2026\.URL[https://openneuro\.org/datasets/ds002799](https://openneuro.org/datasets/ds002799)\.Accessed 2026\-06\-22\.
- \(74\)OpenNeuro Dataset ds002872\.Openneuro dataset ds002872 \(aging\), 2026\.URL[https://openneuro\.org/datasets/ds002872](https://openneuro.org/datasets/ds002872)\.Accessed 2026\-07\-23\.
- \(75\)OpenNeuro Dataset ds002896\.Openneuro dataset ds002896 \(tinnitus\), 2026\.URL[https://openneuro\.org/datasets/ds002896](https://openneuro.org/datasets/ds002896)\.Accessed 2026\-06\-22\.
- \(76\)OpenNeuro Dataset ds002898\.Openneuro dataset ds002898 \(monash RsPETMR\), 2026\.URL[https://openneuro\.org/datasets/ds002898](https://openneuro.org/datasets/ds002898)\.Accessed 2026\-06\-03\.
- \(77\)OpenNeuro Dataset ds002940\.Openneuro dataset ds002940 \(soccercan\), 2026\.URL[https://openneuro\.org/datasets/ds002940](https://openneuro.org/datasets/ds002940)\.Accessed 2026\-07\-23\.
- \(78\)OpenNeuro Dataset ds003007\.Openneuro dataset ds003007 \(two sessions of resting state with closed eyes for patients with depression in treatment course \(nfb, cbt or no treatment groups\)\), 2026\.URL[https://openneuro\.org/datasets/ds003007](https://openneuro.org/datasets/ds003007)\.Accessed 2026\-07\-23\.
- \(79\)OpenNeuro Dataset ds003011\.Openneuro dataset ds003011 \(social processes initiative in neurobiology of the schizophrenia\(s\) traveling human phantoms\), 2026\.URL[https://openneuro\.org/datasets/ds003011](https://openneuro.org/datasets/ds003011)\.Accessed 2026\-07\-23\.
- \(80\)OpenNeuro Dataset ds003027\.Openneuro dataset ds003027 \(plp nf1\), 2026\.URL[https://openneuro\.org/datasets/ds003027](https://openneuro.org/datasets/ds003027)\.Accessed 2026\-06\-22\.
- \(81\)OpenNeuro Dataset ds003037\.Openneuro dataset ds003037 \(sudmex\-tms\), 2026\.URL[https://openneuro\.org/datasets/ds003037](https://openneuro.org/datasets/ds003037)\.Accessed 2026\-06\-22\.
- \(82\)OpenNeuro Dataset ds003085\.Openneuro dataset ds003085 \(temporal dynamics of emotional music\), 2026\.URL[https://openneuro\.org/datasets/ds003085](https://openneuro.org/datasets/ds003085)\.Accessed 2026\-06\-22\.
- \(83\)OpenNeuro Dataset ds003114\.Openneuro dataset ds003114 \(how ovarian hormones influence the behaviroal activation and inhibition system through the dopamine pathway\), 2026\.URL[https://openneuro\.org/datasets/ds003114](https://openneuro.org/datasets/ds003114)\.Associated publication DOI: 10\.1371/journal\.pone\.0237032; accessed 2026\-07\-23\.
- \(84\)OpenNeuro Dataset ds003171\.Openneuro dataset ds003171 \(using anesthesia\-induced loss of consciousness to identify biomarkers of conscious awareness in the healthy human brain\), 2026\.URL[https://openneuro\.org/datasets/ds003171](https://openneuro.org/datasets/ds003171)\.Accessed 2026\-07\-23\.
- \(85\)OpenNeuro Dataset ds003357\.Openneuro dataset ds003357 \(HRV biofeedback\), 2026\.URL[https://openneuro\.org/datasets/ds003357](https://openneuro.org/datasets/ds003357)\.Accessed 2026\-06\-03\.
- \(86\)OpenNeuro Dataset ds003382\.Openneuro dataset ds003382 \(monash vis\-fpet\-fmri\), 2026\.URL[https://openneuro\.org/datasets/ds003382](https://openneuro.org/datasets/ds003382)\.Associated publication DOI: 10\.1016/j\.neuroimage\.2019\.01\.003; accessed 2026\-07\-23\.
- \(87\)OpenNeuro Dataset ds003404\.Openneuro dataset ds003404 \(brain network mechanisms of visual shape completion\), 2026\.URL[https://openneuro\.org/datasets/ds003404](https://openneuro.org/datasets/ds003404)\.Accessed 2026\-07\-23\.
- \(88\)OpenNeuro Dataset ds003424\.Openneuro dataset ds003424 \(identification of an amygdala\-thalamic circuit that acts as a central gain mechanism in taste perception\), 2026\.URL[https://openneuro\.org/datasets/ds003424](https://openneuro.org/datasets/ds003424)\.Associated publication DOI: 10\.1523/JNEUROSCI\.2618\-19\.2020; accessed 2026\-07\-23\.
- \(89\)OpenNeuro Dataset ds003437\.Openneuro dataset ds003437 \(protecting the aging brain \- diet\-study\), 2026\.URL[https://openneuro\.org/datasets/ds003437](https://openneuro.org/datasets/ds003437)\.Associated publication DOI: 10\.1073/pnas\.1913042117; accessed 2026\-07\-23\.
- \(90\)OpenNeuro Dataset ds003455\.Openneuro dataset ds003455 \(protecting the aging brain, case\-study\), 2026\.URL[https://openneuro\.org/datasets/ds003455](https://openneuro.org/datasets/ds003455)\.Associated publication DOI: 10\.1073/pnas\.1913042117; accessed 2026\-07\-23\.
- \(91\)OpenNeuro Dataset ds003468\.Openneuro dataset ds003468 \(iqsec2\), 2026\.URL[https://openneuro\.org/datasets/ds003468](https://openneuro.org/datasets/ds003468)\.Accessed 2026\-06\-22\.
- \(92\)OpenNeuro Dataset ds003469\.Openneuro dataset ds003469 \(speech disfluencies\), 2026\.URL[https://openneuro\.org/datasets/ds003469](https://openneuro.org/datasets/ds003469)\.Accessed 2026\-06\-22\.
- \(93\)OpenNeuro Dataset ds003481\.Openneuro dataset ds003481 \(pragmatic language\), 2026\.URL[https://openneuro\.org/datasets/ds003481](https://openneuro.org/datasets/ds003481)\.Accessed 2026\-06\-22\.
- \(94\)OpenNeuro Dataset ds003540\.Openneuro dataset ds003540 \(multiband acceleration\), 2026\.URL[https://openneuro\.org/datasets/ds003540](https://openneuro.org/datasets/ds003540)\.Accessed 2026\-06\-03\.
- \(95\)OpenNeuro Dataset ds003548\.Openneuro dataset ds003548 \(emotion category and face perception task optimized for multivariate pattern analysis\), 2026\.URL[https://openneuro\.org/datasets/ds003548](https://openneuro.org/datasets/ds003548)\.Associated publication DOI: 10\.1101/2021\.04\.04\.438425; accessed 2026\-07\-23\.
- \(96\)OpenNeuro Dataset ds003574\.Openneuro dataset ds003574 \(reward biases spontaneous neural reactivation during sleep\), 2026\.URL[https://openneuro\.org/datasets/ds003574](https://openneuro.org/datasets/ds003574)\.Accessed 2026\-07\-23\.
- \(97\)OpenNeuro Dataset ds003659\.Openneuro dataset ds003659 \(scantrain\), 2026\.URL[https://openneuro\.org/datasets/ds003659](https://openneuro.org/datasets/ds003659)\.Accessed 2026\-07\-23\.
- \(98\)OpenNeuro Dataset ds003814\.Openneuro dataset ds003814 \(parallel systems for social and spatial reasoning\), 2026\.URL[https://openneuro\.org/datasets/ds003814](https://openneuro.org/datasets/ds003814)\.Accessed 2026\-07\-23\.
- \(99\)OpenNeuro Dataset ds003823\.Openneuro dataset ds003823 \(hrv biofeedback\), 2026\.URL[https://openneuro\.org/datasets/ds003823](https://openneuro.org/datasets/ds003823)\.Accessed 2026\-06\-22\.
- \(100\)OpenNeuro Dataset ds003831\.Openneuro dataset ds003831 \(cognitive control theoretic mechanisms\), 2026\.URL[https://openneuro\.org/datasets/ds003831](https://openneuro.org/datasets/ds003831)\.Accessed 2026\-06\-22\.
- \(101\)OpenNeuro Dataset ds003835\.Openneuro dataset ds003835 \(pe\-update\), 2026\.URL[https://openneuro\.org/datasets/ds003835](https://openneuro.org/datasets/ds003835)\.Associated publication DOI: 10\.1101/2020\.09\.29\.319418; accessed 2026\-07\-23\.
- \(102\)OpenNeuro Dataset ds003849\.Openneuro dataset ds003849 \(frontoparietal plasticity\), 2026\.URL[https://openneuro\.org/datasets/ds003849](https://openneuro.org/datasets/ds003849)\.Accessed 2026\-06\-22\.
- \(103\)OpenNeuro Dataset ds003871\.Openneuro dataset ds003871 \(resting\-state for 34 younger and 28 older adults\), 2026\.URL[https://openneuro\.org/datasets/ds003871](https://openneuro.org/datasets/ds003871)\.Associated publication DOI: 10\.1101/2021\.07\.09\.451777; accessed 2026\-07\-23\.
- \(104\)OpenNeuro Dataset ds003974\.Openneuro dataset ds003974 \(reading brain project l1 adults\), 2026\.URL[https://openneuro\.org/datasets/ds003974](https://openneuro.org/datasets/ds003974)\.Accessed 2026\-06\-22\.
- \(105\)OpenNeuro Dataset ds003988\.Openneuro dataset ds003988 \(reading brain project l2 adults\), 2026\.URL[https://openneuro\.org/datasets/ds003988](https://openneuro.org/datasets/ds003988)\.Accessed 2026\-06\-22\.
- \(106\)OpenNeuro Dataset ds003999\.Openneuro dataset ds003999 \(pre\-post rehabilitation fmri data of post\-stroke patients\.\), 2026\.URL[https://openneuro\.org/datasets/ds003999](https://openneuro.org/datasets/ds003999)\.Accessed 2026\-07\-23\.
- \(107\)OpenNeuro Dataset ds004021\.Openneuro dataset ds004021 \(visuomotor rotation adaptation experiment\), 2026\.URL[https://openneuro\.org/datasets/ds004021](https://openneuro.org/datasets/ds004021)\.Accessed 2026\-06\-22\.
- \(108\)OpenNeuro Dataset ds004101\.Openneuro dataset ds004101 \(infra\-low frequency neurofeedback\), 2026\.URL[https://openneuro\.org/datasets/ds004101](https://openneuro.org/datasets/ds004101)\.Accessed 2026\-06\-22\.
- \(109\)OpenNeuro Dataset ds004109\.Openneuro dataset ds004109 \(emotional learning rest\), 2026\.URL[https://openneuro\.org/datasets/ds004109](https://openneuro.org/datasets/ds004109)\.Accessed 2026\-06\-03\.
- \(110\)OpenNeuro Dataset ds004134\.Openneuro dataset ds004134 \(experience sampling\), 2026\.URL[https://openneuro\.org/datasets/ds004134](https://openneuro.org/datasets/ds004134)\.Accessed 2026\-06\-22\.
- \(111\)OpenNeuro Dataset ds004144\.Openneuro dataset ds004144 \(fibromyalgia emotion regulation dataset\), 2026\.URL[https://openneuro\.org/datasets/ds004144](https://openneuro.org/datasets/ds004144)\.Accessed 2026\-06\-22\.
- \(112\)OpenNeuro Dataset ds004158\.Openneuro dataset ds004158 \(rest\_eye\), 2026\.URL[https://openneuro\.org/datasets/ds004158](https://openneuro.org/datasets/ds004158)\.Accessed 2026\-07\-23\.
- \(113\)OpenNeuro Dataset ds004299\.Openneuro dataset ds004299 \(habit learning\), 2026\.URL[https://openneuro\.org/datasets/ds004299](https://openneuro.org/datasets/ds004299)\.Accessed 2026\-06\-22\.
- \(114\)OpenNeuro Dataset ds004469\.Openneuro dataset ds004469 \(temporal lobe epilepsy \- unam\), 2026\.URL[https://openneuro\.org/datasets/ds004469](https://openneuro.org/datasets/ds004469)\.Accessed 2026\-06\-22\.
- \(115\)OpenNeuro Dataset ds004498\.Openneuro dataset ds004498 \(perinatal stroke\), 2026\.URL[https://openneuro\.org/datasets/ds004498](https://openneuro.org/datasets/ds004498)\.Associated publication DOI: 10\.1101/2022\.10\.26\.513940; 10\.1093/cercor/bhab387; accessed 2026\-07\-23\.
- \(116\)OpenNeuro Dataset ds004533\.Openneuro dataset ds004533 \(ironia vev\), 2026\.URL[https://openneuro\.org/datasets/ds004533](https://openneuro.org/datasets/ds004533)\.Accessed 2026\-07\-23\.
- \(117\)OpenNeuro Dataset ds004627\.Openneuro dataset ds004627 \(nimh cat\-d\), 2026\.URL[https://openneuro\.org/datasets/ds004627](https://openneuro.org/datasets/ds004627)\.Accessed 2026\-06\-22\.
- \(118\)OpenNeuro Dataset ds004630\.Openneuro dataset ds004630 \(spida\-mri\), 2026\.URL[https://openneuro\.org/datasets/ds004630](https://openneuro.org/datasets/ds004630)\.Accessed 2026\-06\-22\.
- \(119\)OpenNeuro Dataset ds004645\.Openneuro dataset ds004645 \(hemodynamic timing\), 2026\.URL[https://openneuro\.org/datasets/ds004645](https://openneuro.org/datasets/ds004645)\.Accessed 2026\-06\-03\.
- \(120\)OpenNeuro Dataset ds004648\.Openneuro dataset ds004648 \(gut\-brain axis\), 2026\.URL[https://openneuro\.org/datasets/ds004648](https://openneuro.org/datasets/ds004648)\.Accessed 2026\-06\-03\.
- \(121\)OpenNeuro Dataset ds004697\.Openneuro dataset ds004697 \(food brain study\), 2026\.URL[https://openneuro\.org/datasets/ds004697](https://openneuro.org/datasets/ds004697)\.Accessed 2026\-06\-22\.
- \(122\)OpenNeuro Dataset ds004712\.Openneuro dataset ds004712 \(on\-harmony\), 2026\.URL[https://openneuro\.org/datasets/ds004712](https://openneuro.org/datasets/ds004712)\.Accessed 2026\-06\-22\.
- \(123\)OpenNeuro Dataset ds004725\.Openneuro dataset ds004725 \(oxytocin\), 2026\.URL[https://openneuro\.org/datasets/ds004725](https://openneuro.org/datasets/ds004725)\.Accessed 2026\-06\-22\.
- \(124\)OpenNeuro Dataset ds004737\.Openneuro dataset ds004737 \(cs\-dsi\), 2026\.URL[https://openneuro\.org/datasets/ds004737](https://openneuro.org/datasets/ds004737)\.Associated publication DOI: 10\.1101/2023\.02\.22\.529546; accessed 2026\-07\-23\.
- \(125\)OpenNeuro Dataset ds004765\.Openneuro dataset ds004765 \(language fmri\), 2026\.URL[https://openneuro\.org/datasets/ds004765](https://openneuro.org/datasets/ds004765)\.Accessed 2026\-07\-23\.
- \(126\)OpenNeuro Dataset ds004787\.Openneuro dataset ds004787 \(nimh meter\), 2026\.URL[https://openneuro\.org/datasets/ds004787](https://openneuro.org/datasets/ds004787)\.Accessed 2026\-06\-22\.
- \(127\)OpenNeuro Dataset ds004884\.Openneuro dataset ds004884 \(aphasia recovery cohort\), 2026\.URL[https://openneuro\.org/datasets/ds004884](https://openneuro.org/datasets/ds004884)\.Accessed 2026\-06\-22\.
- \(128\)OpenNeuro Dataset ds004965\.Openneuro dataset ds004965 \(truecrime\), 2026\.URL[https://openneuro\.org/datasets/ds004965](https://openneuro.org/datasets/ds004965)\.Accessed 2026\-06\-22\.
- \(129\)OpenNeuro Dataset ds005003\.Retrospective task and rest fmri data from brain tumor patients, 2026\.URL[https://openneuro\.org/datasets/ds005003](https://openneuro.org/datasets/ds005003)\.Accessed 2026\-05\-12\.
- \(130\)OpenNeuro Dataset ds005016\.Openneuro dataset ds005016 \(mbsr\), 2026\.URL[https://openneuro\.org/datasets/ds005016](https://openneuro.org/datasets/ds005016)\.Accessed 2026\-06\-22\.
- \(131\)OpenNeuro Dataset ds005017\.Openneuro dataset ds005017 \(pop\), 2026\.URL[https://openneuro\.org/datasets/ds005017](https://openneuro.org/datasets/ds005017)\.Accessed 2026\-07\-23\.
- \(132\)OpenNeuro Dataset ds005026\.Openneuro dataset ds005026 \(hearing loss connectome\), 2026\.URL[https://openneuro\.org/datasets/ds005026](https://openneuro.org/datasets/ds005026)\.Accessed 2026\-06\-22\.
- \(133\)OpenNeuro Dataset ds005073\.Openneuro dataset ds005073 \(the neural basis of visual shape completion in schizophrenia and bipolar disorder\), 2026\.URL[https://openneuro\.org/datasets/ds005073](https://openneuro.org/datasets/ds005073)\.Accessed 2026\-07\-23\.
- \(134\)OpenNeuro Dataset ds005088\.Openneuro dataset ds005088 \(neural correlates of lidocaine analgesic \(nla\) study\), 2026\.URL[https://openneuro\.org/datasets/ds005088](https://openneuro.org/datasets/ds005088)\.Accessed 2026\-07\-23\.
- \(135\)OpenNeuro Dataset ds005115\.Openneuro dataset ds005115 \(28andhe\), 2026\.URL[https://openneuro\.org/datasets/ds005115](https://openneuro.org/datasets/ds005115)\.Associated publication DOI: 10\.1523/JNEUROSCI\.1856\-23\.2024; 10\.1101/2024\.04\.10\.588906; accessed 2026\-07\-23\.
- \(136\)OpenNeuro Dataset ds005118\.Openneuro dataset ds005118 \(weill cornell medicine multi\-echo \(wcm\-me\) dataset\), 2026\.URL[https://openneuro\.org/datasets/ds005118](https://openneuro.org/datasets/ds005118)\.Accessed 2026\-07\-23\.
- \(137\)OpenNeuro Dataset ds005127\.Openneuro dataset ds005127 \(amri 16\-n\-0031 sleep1\), 2026\.URL[https://openneuro\.org/datasets/ds005127](https://openneuro.org/datasets/ds005127)\.Associated publication DOI: 10\.1016/j\.jneumeth\.2018\.09\.019; accessed 2026\-07\-23\.
- \(138\)OpenNeuro Dataset ds005230\.Openneuro dataset ds005230 \(reinforcement\-learning generalization\), 2026\.URL[https://openneuro\.org/datasets/ds005230](https://openneuro.org/datasets/ds005230)\.Accessed 2026\-06\-22\.
- \(139\)OpenNeuro Dataset ds005264\.Openneuro dataset ds005264 \(socal kinesia and incentivization for parkinson’s disease \(skip\): Ultra\-high field functional connectivity\), 2026\.URL[https://openneuro\.org/datasets/ds005264](https://openneuro.org/datasets/ds005264)\.Accessed 2026\-07\-23\.
- \(140\)OpenNeuro Dataset ds005270\.Openneuro dataset ds005270 \(bold variability during cognitive control\), 2026\.URL[https://openneuro\.org/datasets/ds005270](https://openneuro.org/datasets/ds005270)\.Accessed 2026\-06\-22\.
- \(141\)OpenNeuro Dataset ds005339\.Openneuro dataset ds005339 \(yale reading\), 2026\.URL[https://openneuro\.org/datasets/ds005339](https://openneuro.org/datasets/ds005339)\.Accessed 2026\-06\-22\.
- \(142\)OpenNeuro Dataset ds005346\.Openneuro dataset ds005346 \(baba: Naturalistic fmri and meg dataset\), 2026\.URL[https://openneuro\.org/datasets/ds005346](https://openneuro.org/datasets/ds005346)\.Accessed 2026\-07\-23\.
- \(143\)OpenNeuro Dataset ds005405\.Openneuro dataset ds005405 \(protecting the aging brain \- fmri study of the brain in ketosis\), 2026\.URL[https://openneuro\.org/datasets/ds005405](https://openneuro.org/datasets/ds005405)\.Accessed 2026\-07\-23\.
- \(144\)OpenNeuro Dataset ds005418\.Openneuro dataset ds005418 \(CO2 resting\), 2026\.URL[https://openneuro\.org/datasets/ds005418](https://openneuro.org/datasets/ds005418)\.Accessed 2026\-06\-03\.
- \(145\)OpenNeuro Dataset ds005449\.Openneuro dataset ds005449 \(valenced tactile information\), 2026\.URL[https://openneuro\.org/datasets/ds005449](https://openneuro.org/datasets/ds005449)\.Accessed 2026\-06\-22\.
- \(146\)OpenNeuro Dataset ds005464\.Openneuro dataset ds005464 \(priority\), 2026\.URL[https://openneuro\.org/datasets/ds005464](https://openneuro.org/datasets/ds005464)\.Accessed 2026\-07\-23\.
- \(147\)OpenNeuro Dataset ds005525\.Openneuro dataset ds005525 \(circadian misalignment and energy balance\), 2026\.URL[https://openneuro\.org/datasets/ds005525](https://openneuro.org/datasets/ds005525)\.Accessed 2026\-07\-23\.
- \(148\)OpenNeuro Dataset ds005573\.Openneuro dataset ds005573 \(a comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients\), 2026\.URL[https://openneuro\.org/datasets/ds005573](https://openneuro.org/datasets/ds005573)\.Associated publication DOI: 10\.1016/j\.nicl\.2019\.101850; accessed 2026\-07\-23\.
- \(149\)OpenNeuro Dataset ds005598\.Openneuro dataset ds005598 \(visuomotor rotation learning and reward\-based motor learning\), 2026\.URL[https://openneuro\.org/datasets/ds005598](https://openneuro.org/datasets/ds005598)\.Accessed 2026\-06\-22\.
- \(150\)OpenNeuro Dataset ds005603\.Openneuro dataset ds005603 \(mitchell\_hacker\_2013\), 2026\.URL[https://openneuro\.org/datasets/ds005603](https://openneuro.org/datasets/ds005603)\.Associated publication DOI: 10\.1227/NEU\.0000000000000141; accessed 2026\-07\-23\.
- \(151\)OpenNeuro Dataset ds005713\.Openneuro dataset ds005713 \(large\-scale pre/post\-surgical patients\), 2026\.URL[https://openneuro\.org/datasets/ds005713](https://openneuro.org/datasets/ds005713)\.Accessed 2026\-06\-22\.
- \(152\)OpenNeuro Dataset ds005752\.Openneuro dataset ds005752 \(nimh healthy research volunteer dataset\), 2026\.URL[https://openneuro\.org/datasets/ds005752](https://openneuro.org/datasets/ds005752)\.Accessed 2026\-06\-22\.
- \(153\)OpenNeuro Dataset ds005754\.Openneuro dataset ds005754 \(inhibitory control youth\), 2026\.URL[https://openneuro\.org/datasets/ds005754](https://openneuro.org/datasets/ds005754)\.Accessed 2026\-06\-22\.
- \(154\)OpenNeuro Dataset ds005795\.Openneuro dataset ds005795 \(multi\-clarid\), 2026\.URL[https://openneuro\.org/datasets/ds005795](https://openneuro.org/datasets/ds005795)\.Accessed 2026\-06\-22\.
- \(155\)OpenNeuro Dataset ds005849\.Openneuro dataset ds005849 \(the dbs\-fmri dataset\), 2026\.URL[https://openneuro\.org/datasets/ds005849](https://openneuro.org/datasets/ds005849)\.Accessed 2026\-07\-23\.
- \(156\)OpenNeuro Dataset ds005874\.Appetite neural correlates in motor neuron disease, 2026\.URL[https://openneuro\.org/datasets/ds005874](https://openneuro.org/datasets/ds005874)\.Accessed 2026\-05\-12\.
- \(157\)OpenNeuro Dataset ds005875\.Openneuro dataset ds005875 \(china’s social fake news database release with brain structural, functional, and behavioural measures\), 2026\.URL[https://openneuro\.org/datasets/ds005875](https://openneuro.org/datasets/ds005875)\.Accessed 2026\-07\-23\.
- \(158\)OpenNeuro Dataset ds005892\.Openneuro dataset ds005892 \(hc / pd\-nc / pd\-mci resting\-state mri\), 2026\.URL[https://openneuro\.org/datasets/ds005892](https://openneuro.org/datasets/ds005892)\.Accessed 2026\-06\-22\.
- \(159\)OpenNeuro Dataset ds005896\.Adolescent health and development in context \(AHDC\), 2026\.URL[https://openneuro\.org/datasets/ds005896](https://openneuro.org/datasets/ds005896)\.Accessed 2026\-05\-12\.
- \(160\)OpenNeuro Dataset ds005917\.Openneuro dataset ds005917 \(nimh ketamine mechanism of action study\), 2026\.URL[https://openneuro\.org/datasets/ds005917](https://openneuro.org/datasets/ds005917)\.Accessed 2026\-06\-22\.
- \(161\)OpenNeuro Dataset ds006045\.Openneuro dataset ds006045 \(anxiety CBT rest\), 2026\.URL[https://openneuro\.org/datasets/ds006045](https://openneuro.org/datasets/ds006045)\.Accessed 2026\-06\-03\.
- \(162\)OpenNeuro Dataset ds006110\.Openneuro dataset ds006110 \(psiconnect\), 2026\.URL[https://openneuro\.org/datasets/ds006110](https://openneuro.org/datasets/ds006110)\.Accessed 2026\-06\-22\.
- \(163\)OpenNeuro Dataset ds006131\.Openneuro dataset ds006131 \(pafin\), 2026\.URL[https://openneuro\.org/datasets/ds006131](https://openneuro.org/datasets/ds006131)\.Accessed 2026\-06\-22\.
- \(164\)OpenNeuro Dataset ds006148\.Openneuro dataset ds006148 \(qnl negativebold database\), 2026\.URL[https://openneuro\.org/datasets/ds006148](https://openneuro.org/datasets/ds006148)\.Accessed 2026\-07\-23\.
- \(165\)OpenNeuro Dataset ds006303\.Openneuro dataset ds006303 \(pediatric anxiety rest\), 2026\.URL[https://openneuro\.org/datasets/ds006303](https://openneuro.org/datasets/ds006303)\.Accessed 2026\-06\-03\.
- \(166\)OpenNeuro Dataset ds006395\.Openneuro dataset ds006395 \(dystonia rest\), 2026\.URL[https://openneuro\.org/datasets/ds006395](https://openneuro.org/datasets/ds006395)\.Accessed 2026\-06\-22\.
- \(167\)OpenNeuro Dataset ds006515\.Openneuro dataset ds006515 \(chicago attention and thoughts\), 2026\.URL[https://openneuro\.org/datasets/ds006515](https://openneuro.org/datasets/ds006515)\.Accessed 2026\-07\-23\.
- \(168\)OpenNeuro Dataset ds006644\.Openneuro dataset ds006644 \(dmt\-har\-med: Effects of dmt and harmine during meditation\), 2026\.URL[https://openneuro\.org/datasets/ds006644](https://openneuro.org/datasets/ds006644)\.Accessed 2026\-06\-22\.
- \(169\)OpenNeuro Dataset ds006683\.Openneuro dataset ds006683 \(rt disengagement\), 2026\.URL[https://openneuro\.org/datasets/ds006683](https://openneuro.org/datasets/ds006683)\.Accessed 2026\-06\-22\.
- \(170\)OpenNeuro Dataset ds006686\.Openneuro dataset ds006686 \(prototype/exemplar aging\), 2026\.URL[https://openneuro\.org/datasets/ds006686](https://openneuro.org/datasets/ds006686)\.Accessed 2026\-06\-03\.
- \(171\)OpenNeuro Dataset ds006693\.Openneuro dataset ds006693 \(nodeap\), 2026\.URL[https://openneuro\.org/datasets/ds006693](https://openneuro.org/datasets/ds006693)\.Accessed 2026\-06\-22\.
- \(172\)OpenNeuro Dataset ds006707\.Openneuro dataset ds006707 \(night owls scan club\), 2026\.URL[https://openneuro\.org/datasets/ds006707](https://openneuro.org/datasets/ds006707)\.Accessed 2026\-07\-23\.
- \(173\)OpenNeuro Dataset ds006893\.Openneuro dataset ds006893 \(insulin menstrual rest\), 2026\.URL[https://openneuro\.org/datasets/ds006893](https://openneuro.org/datasets/ds006893)\.Accessed 2026\-06\-03\.
- \(174\)OpenNeuro Dataset ds006948\.Openneuro dataset ds006948 \(linking subjective experience of anxiety to brain function using natural language processing\.\), 2026\.URL[https://openneuro\.org/datasets/ds006948](https://openneuro.org/datasets/ds006948)\.Associated publication DOI: 10\.1093/scan/nsaf099; accessed 2026\-07\-23\.
- \(175\)OpenNeuro Dataset ds006952\.Openneuro dataset ds006952 \(psych\-rest\), 2026\.URL[https://openneuro\.org/datasets/ds006952](https://openneuro.org/datasets/ds006952)\.Accessed 2026\-06\-22\.
- \(176\)OpenNeuro Dataset ds007036\.Openneuro dataset ds007036 \(ALS\-FTD UHF\), 2026\.URL[https://openneuro\.org/datasets/ds007036](https://openneuro.org/datasets/ds007036)\.Accessed 2026\-06\-03\.
- \(177\)OpenNeuro Dataset ds007063\.Openneuro dataset ds007063 \(hippocampal memory rest\), 2026\.URL[https://openneuro\.org/datasets/ds007063](https://openneuro.org/datasets/ds007063)\.Accessed 2026\-06\-03\.
- \(178\)OpenNeuro Dataset ds007116\.Openneuro dataset ds007116 \(penn LEAD\), 2026\.URL[https://openneuro\.org/datasets/ds007116](https://openneuro.org/datasets/ds007116)\.Accessed 2026\-06\-03\.
- \(179\)OpenNeuro Dataset ds007286\.Openneuro dataset ds007286 \(yale neuroconnect\), 2026\.URL[https://openneuro\.org/datasets/ds007286](https://openneuro.org/datasets/ds007286)\.Accessed 2026\-06\-22\.
- \(180\)OpenNeuro Dataset ds007328\.Openneuro dataset ds007328 \(dense longitudinal single\-subject multimodal mri dataset acquired via self\-administered scanning\), 2026\.URL[https://openneuro\.org/datasets/ds007328](https://openneuro.org/datasets/ds007328)\.Accessed 2026\-07\-23\.
- \(181\)OpenNeuro Dataset ds007345\.Openneuro dataset ds007345 \(olfactory meningioma\), 2026\.URL[https://openneuro\.org/datasets/ds007345](https://openneuro.org/datasets/ds007345)\.Accessed 2026\-06\-03\.
- \(182\)OpenNeuro Dataset ds007354\.Openneuro dataset ds007354 \(5\.0t visual scene\), 2026\.URL[https://openneuro\.org/datasets/ds007354](https://openneuro.org/datasets/ds007354)\.Accessed 2026\-06\-22\.
- \(183\)OpenNeuro Dataset ds007522\.Openneuro dataset ds007522 \(precision aging network\), 2026\.URL[https://openneuro\.org/datasets/ds007522](https://openneuro.org/datasets/ds007522)\.Accessed 2026\-06\-22\.
- \(184\)OpenNeuro Dataset ds007694\.Openneuro dataset ds007694 \(GRACE\), 2026\.URL[https://openneuro\.org/datasets/ds007694](https://openneuro.org/datasets/ds007694)\.Accessed 2026\-06\-03\.
- \(185\)OpenNeuro Dataset ds007908\.Openneuro dataset ds007908 \(multi\-scale, multi\-modal imaging assessment of trajectories of cognitive impairment in multiple sclerosis\), 2026\.URL[https://openneuro\.org/datasets/ds007908](https://openneuro.org/datasets/ds007908)\.Accessed 2026\-07\-23\.
- \(186\)Josue Ortega Caro, Antonio Henrique de Oliveira Fonseca, Syed Rizvi, Matteo Rosati, Christopher Averill, James Cross, Prateek Mittal, Emanuele Zappala, Rahul Dhodapkar, Chadi Abdallah, and David van Dijk\.BrainLM: A foundation model for brain activity recordings\.In*International Conference on Learning Representations*, 2024\.URL[https://proceedings\.iclr\.cc/paper\_files/paper/2024/hash/029ce70401321de3808b3ac39e1ab167\-Abstract\-Conference\.html](https://proceedings.iclr.cc/paper_files/paper/2024/hash/029ce70401321de3808b3ac39e1ab167-Abstract-Conference.html)\.
- \(187\)Russell A\. Poldrack, Timothy O\. Laumann, Oluwasanmi Koyejo, Brandon Gregory, Amber Hover, M\. Yvonne Chen, Krzysztof J\. Gorgolewski, Jonathan Luci, Sungjae J\. Joo, Ryan L\. Boyd, Steven Hunicke\-Smith, Zack B\. Simpson, Timothy Caven, Vanessa Sochat, Jarrod M\. Shine, Evan Gordon, Abraham Z\. Snyder, Babatunde Adeyemo, Steven E\. Petersen, and Jeanette A\. Mumford\.Long\-term neural and physiological phenotyping of a single human\.*Nature Communications*, 6:8885, 2015\.doi:10\.1038/ncomms9885\.URL[https://www\.nature\.com/articles/ncomms9885](https://www.nature.com/articles/ncomms9885)\.
- \(188\)Russell A\. Poldrack, Eliza Congdon, William Triplett, et al\.A phenome\-wide examination of neural and cognitive function\.*Scientific Data*, 3\(1\):160110, 2016\.doi:10\.1038/sdata\.2016\.110\.
- \(189\)Chris Racey, Christina Kampoureli, Oscar Bowen\-Hill, Mathilde Bauer, Ivor Simpson, Charlotte Rae, Magda del Rio, Julia Simner, and Jamie Ward\.An open science mri database of over 100 synaesthetic brains and accompanying deep phenotypic information\.*Scientific Data*, 10\(1\):766, 2023\.doi:10\.1038/s41597\-023\-02664\-4\.URL[https://www\.nature\.com/articles/s41597\-023\-02664\-4](https://www.nature.com/articles/s41597-023-02664-4)\.
- \(190\)Giovanni A\. Salum, Carina de Giusti, Laila Souza, et al\.Cohort profile: Brazilian high\-risk cohort for mental health conditions \(BHRC\)\.*International Journal of Epidemiology*, 54\(6\), 2025\.doi:10\.1093/ije/dyaf192\.
- \(191\)Theodore D\. Satterthwaite, Mark A\. Elliott, Kosha Ruparel, et al\.Neuroimaging of the philadelphia neurodevelopmental cohort\.*NeuroImage*, 86:544–553, 2014\.doi:10\.1016/j\.neuroimage\.2013\.07\.064\.
- \(192\)Lukas Snoek, Maite M\. van der Miesen, Tinka Beemsterboer, et al\.The amsterdam open MRI collection, a set of multimodal MRI datasets for individual difference analyses\.*Scientific Data*, 8\(1\):85, 2021\.doi:10\.1038/s41597\-021\-00870\-6\.
- \(193\)R\. Nathan Spreng, Eva Dimas, Luhang Mwilambwe\-Tshilobo, et al\.Neurocognitive aging data release with behavioral, structural and multi\-echo functional mri measures\.*Scientific Data*, 9\(1\):146, 2022\.doi:10\.1038/s41597\-022\-01231\-7\.URL[https://www\.nature\.com/articles/s41597\-022\-01231\-7](https://www.nature.com/articles/s41597-022-01231-7)\.
- \(194\)Lachlan T\. Strike, Greig I\. de Zubicaray, Katie L\. McMahon, Paul M\. Thompson, Nicholas G\. Martin, and Margaret J\. Wright\.Queensland twin adolescent brain \(QTAB\), 2023\.URL[https://openneuro\.org/datasets/ds004146](https://openneuro.org/datasets/ds004146)\.
- \(195\)Sho Tanaka, Masaaki Ide, Masaki Miyakoshi, et al\.A multi\-site, multi\-disorder resting\-state magnetic resonance image database\.*Scientific Data*, 8\(1\):227, 2021\.doi:10\.1038/s41597\-021\-01004\-8\.URL[https://www\.nature\.com/articles/s41597\-021\-01004\-8](https://www.nature.com/articles/s41597-021-01004-8)\.
- \(196\)The ADHD\-200 Consortium\.The ADHD\-200 consortium: a model to advance the translational potential of neuroimaging in clinical neuroscience\.*Frontiers in Systems Neuroscience*, 6:62, 2012\.doi:10\.3389/fnsys\.2012\.00062\.
- \(197\)David C\. Van Essen, Stephen M\. Smith, Deanna M\. Barch, Timothy E\. J\. Behrens, Essa Yacoub, and Kamil Ugurbil\.The WU\-minn human connectome project: An overview\.*NeuroImage*, 80:62–79, 2013\.doi:10\.1016/j\.neuroimage\.2013\.05\.041\.
- \(198\)Dongtao Wei, Kaixiang Zhuang, Lei Ai, et al\.Structural and functional brain scans from the cross\-sectional southwest university adult lifespan dataset\.*Scientific Data*, 5\(1\):180134, 2018\.doi:10\.1038/sdata\.2018\.134\.
- \(199\)Xinxu Wei, Kanhao Zhao, Yong Jiao, Lifang He, and Yu Zhang\.A brain graph foundation model: Pre\-training and prompt\-tuning across broad atlases and disorders\.In*International Conference on Learning Representations*, 2026\.doi:10\.48550/arXiv\.2506\.02044\.URL[https://arxiv\.org/abs/2506\.02044](https://arxiv.org/abs/2506.02044)\.
- \(200\)Tao Wu, Jie Wang, Cheng Wang, Mark Hallett, Yufeng Zang, and Xi\-Nian Wu\.Exploring the reproducibility of functional connectivity alterations in parkinson’s disease\.*PLOS ONE*, 12\(11\):e0188196, 2017\.doi:10\.1371/journal\.pone\.0188196\.URL[https://journals\.plos\.org/plosone/article?id=10\.1371/journal\.pone\.0188196](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0188196)\.
- \(201\)Chao\-Gan Yan, Xiao Chen, Le Li, et al\.Reduced default mode network functional connectivity in patients with recurrent major depressive disorder\.*Proceedings of the National Academy of Sciences*, 116\(18\):9078–9083, 2019\.doi:10\.1073/pnas\.1900390116\.
- \(202\)Yanwu Yang, Chenfei Ye, Guinan Su, Ziyao Zhang, Zhikai Chang, Hairui Chen, Piu Chan, Yue Yu, and Ting Ma\.BrainMass: Advancing brain network analysis for diagnosis with large\-scale self\-supervised learning\.*arXiv preprint arXiv:2403\.01433*, 2024\.doi:10\.48550/arXiv\.2403\.01433\.URL[https://arxiv\.org/abs/2403\.01433](https://arxiv.org/abs/2403.01433)\.
- \(203\)Jiahe Zhang, Daniel Chen, Paige Deming, et al\.Cortical and subcortical mapping of the human allostatic–interoceptive system using 7 tesla fmri\.*Nature Neuroscience*, 28:2380–2391, 2025\.doi:10\.1038/s41593\-025\-02087\-x\.URL[https://www\.nature\.com/articles/s41593\-025\-02087\-x](https://www.nature.com/articles/s41593-025-02087-x)\.
- \(204\)Xi\-Nian Zuo, Jeffrey S\. Anderson, Pierre Bellec, et al\.An open science resource for establishing reliability and reproducibility in functional connectomics\.*Scientific Data*, 1\(1\):140049, 2014\.doi:10\.1038/sdata\.2014\.49\.