MindScale Sample Data

Questionnaire-based mental health screening and symptom assessment

Overview · Contents · Data Format · Using with MindScale · Usage Notes

📋 Overview

This directory contains 100 questionnaire-format sample records for inspecting MindScale's data structure and preparing input demonstrations. Each JSON file represents one sample.

Screening: 50 samples covering anxiety, depression, life satisfaction, and life stress.
Assessment: 50 samples covering five questionnaire-derived symptom domains, with target categories defined below.
Scope: These examples are a small sample collection, not the full study cohort. They are not sufficient to reproduce the paper's population-scale performance estimates.

⬆️ Back to top

📂 Contents

open_data/
|-- README.md
|-- At_risk_samples/       50 screening JSON files
`-- Assessment_samples/    50 assessment JSON files

Both folders include primary- and secondary-school questionnaire formats. The number and names of questionnaire items therefore vary by school stage. Sample manifests and sampling-weight files are not included in this directory.

⬆️ Back to top

🧾 Data Format

All sample files use UTF-8 JSON encoding.

Field	Contents
`input`	Observed background and environmental questionnaire responses as key-value pairs
`mask`	Item response records, each containing `value` and an ordered `candidate` list
`label`	Four screening targets in `At_risk_samples`
`re_label`	Five assessment targets in `Assessment_samples`; some screening files also contain ancillary assessment results here
`anxiety_depression`	Auxiliary anxiety and depression outcomes in assessment samples

Item Response Records

The following is an illustrative entry within mask:

{
  "Self-strength": {
    "value": "Basically agree",
    "candidate": [
      "Strongly disagree",
      "Not quite agree",
      "Basically agree",
      "Strongly agree"
    ]
  }
}

Meaning of mask: This field contains response values, not only missing items. Pretraining predicts these responses, while downstream input builders incorporate their values as questionnaire features.
Missing responses: Some assessment item records use "-" for an unavailable response. Preserve this marker and the original candidate list.
Auxiliary outcomes: label and re_label are stage-specific fields, not interchangeable aliases. Ancillary outcomes should be kept separate from the selected prediction targets.

Screening Targets

Read the screening target mapping from label.

Target key	Candidate labels, in model order
`anxiety`	`No anxiety`, `Mild anxiety`, `Severe anxiety`
`depression`	`No depression`, `Mild depression`, `Severe depression`
`satisfaction`	`Low life satisfaction`, `High life satisfaction`
`stress`	`No life stress`, `Low life stress`, `High life stress`

Assessment Targets

Read the assessment target mapping from re_label.

Target key	Candidate labels, in model order
`re_suicidal`	`Mild suicidal ideation`, `Severe suicidal ideation`
`re_obsessive`	`Mild obsessive symptoms`, `Severe obsessive symptoms`
`re_psychotic`	`Mild psychotic symptoms`, `Severe psychotic symptoms`
`re_mania`	`Mild mania`, `Severe mania`
`re_ADHD`	`Mild ADHD`, `Severe ADHD`

These strings are the model's questionnaire-derived categories, not clinical diagnoses. Preserve their spelling and capitalization when using released adapters; the corresponding mappings are defined in the code's dataset_info.py.

⬆️ Back to top

🚀 Using with MindScale

Select the stage

Use At_risk_samples for --task_type Screen and Assessment_samples for --task_type Assessment. The screening samples' input and mask fields can also supply pretraining inputs; screening outcomes are not pretraining targets.
Prepare the target mapping for the current input builder

The comparative builder reads targets from label, whereas the non-comparative builder reads them from re_label. Prepare working copies with the selected stage's target mapping in the required field:

Stage	Source target field	Comparative mode	Non-comparative mode
`Screen`	`label`	Keep screening targets in `label`	Copy screening targets to `re_label` in a working copy
`Assessment`	`re_label`	Copy assessment targets to `label` in a working copy	Keep assessment targets in `re_label`

Preserve any original auxiliary outcomes separately when preparing these copies.

Prepare manifests

For downstream runs, create the following UTF-8 text files. Sample paths must be relative to --data_path, with one entry per line.

train.txt          Reference/training sample paths
grade_train.txt    School-stage label for each reference/training sample
mean_IP_train.txt  Template-sampling weight for each reference/training sample
test.txt           Test sample paths
grade_test.txt     School-stage label for each test sample

The three training files must be line-aligned with one another; the two test files must be line-aligned with one another. Use separate reference and test samples, and calculate sampling weights from reference/training labels only.

Map school stages

The grade manifests use the following labels, derived from input["School-status"]:

Sample value	Grade-manifest label
`Schoolchild`	`Primary school`
`Middle school student`	`Middle school`
`High school student`	`Middle school`

With two templates, each school-stage group used during training needs at least three reference samples; evaluation needs at least two reference samples. Templates must be drawn from the reference/training set.

Load model weights

Follow the accompanying MindScale code README for installation, checkpoint loading, and prediction export. Three-stage LoRA adapters are available on Hugging Face.

⬆️ Back to top

📝 Usage Notes

Research scope: Use these samples for authorized research, schema inspection, and integration checks. They are not a clinical assessment service.
Privacy: Questionnaire records contain sensitive background and mental health information. Protect participant privacy and do not attempt to re-identify individuals or link records to identifiable people.
Release permissions: Availability in this directory does not by itself establish permission for unrestricted reuse or redistribution. Follow the data provider's applicable authorization and data-use terms.

⬆️ Back to top