DATASHEET.md

# Datasheet: STAQ (Synthetic Technology Assistance Queries)

This datasheet follows the "Datasheets for Datasets" framework (Gebru et al.,
2021). For dataset contents and usage, see the [README](README.md); for the
language variety and population framing, see the [data statement](DATA_STATEMENT.md).

## Motivation

### For what purpose was the dataset created?

STAQ was created to support building age-inclusive artificial intelligence (AI)
systems. Large-scale training data underrepresents older adults, and asking
older adults to generate the volume of data needed to train or fine-tune
foundation models is impractical and ethically burdensome. STAQ provides a
scalable, synthetic bridge for model training and evaluation, grounded in real
observed behavior rather than stereotypes. It is meant to complement, not
replace, participatory design and human evaluation.

### Who created the dataset?

The dataset was created by Hasti Sharifi, Homaira Huda Shomee, Melissa Lamar,
Sourav Medya, and Debaleena Chattopadhyay, in connection with the ASSETS '26
paper "Helping the Helper: LLM-Assisted Problem Articulation for Older Adults
Seeking Technology Support."

### Who funded the creation of the dataset?

This work was supported in part by the National Institute on Aging of the
National Institutes of Health under Award No. P30AG083255. The content is
solely the responsibility of the authors and does not necessarily represent the
official views of the National Institutes of Health.

## Composition

### What do the instances represent?

Each instance is a technology help-seeking query written in the communication
style of older adults, paired with a paraphrase that clarifies
the intended meaning, a communication-characteristic label, and (for a sampled
subset) an expert plausibility rating.

### How many instances are there?

There are 514 instances, distributed across four categories, as shown below:

| Category | Count |
| --- | --- |
| under-specification | 154 |
| verbosity | 120 |
| over-specification | 120 |
| incompleteness | 120 |

### Is the dataset a sample, or does it contain all instances?

The dataset is a generated set, not a sample of a larger fixed collection. It was
produced to cover the four target communication characteristics; the counts
above reflect the released composition.

### What data does each instance consist of?

Each row has five fields: `id` (unique identifier), `query` (older-adult-style
query), `rephrased_query` (AI-generated paraphrase), `category` (one of the four
characteristics), and `expert_rating` (expert plausibility rating).

### Is there a label or target associated with each instance?

Yes. `category` labels every instance. `rephrased_query` serves as a paired
reference/target for paraphrasing tasks. `expert_rating` is an additional label
available only for the evaluated subset.

### Is any information missing from individual instances?

Yes. `expert_rating` is populated only for the 50-query face-validity sample; the
remaining rows are unrated. Missing ratings should not be read as low ratings.

### Are relationships between instances made explicit?

Each `query` is explicitly paired with its `rephrased_query`, and grouped by
`category`. There are no other cross-instance relationships.

### Are there recommended data splits?

No predefined train/validation/test splits are provided; the dataset is a single
flat file. Users may split it as appropriate, ideally stratified by `category`.

### Are there errors, sources of noise, or redundancies?

Because the data is model-generated, some queries are less realistic than others:
in the face-validity review, 9 of 50 sampled queries were rated "unlikely" and 1
"possibly." Users should expect some variation in realism.

### Is the dataset self-contained?

Yes. The released dataset is a single self-contained CSV file. It was grounded in
a private diary-study corpus (see Collection process), which is not included.

### Does the dataset contain confidential or sensitive data?

No. STAQ is synthetic and contains no real personal information. Scenarios
reference common platforms and apps (for example, email, video calls, ride-hailing)
but describe fictional situations and no real individuals. The dataset concerns a
population defined by age but contains no real individuals' data.

## Collection process

### How was the data acquired?

The queries were generated by a large language model (GPT-4o) using few-shot
prompting, anchored in real, unstructured technology-support queries from a
formative diary study with older adults.

### What mechanisms or procedures were used?

Generation used few-shot prompting: each prompt contained three example pairs (an
original older-adult-style query and its expert paraphrase) and instructed the
model to produce one new example in the same conversational style, output as
JSON. The model was directed to vary the technical issues and was not asked to
emulate a generic "older adult." The full prompt appears in the paper's appendix.

### Who was involved in the collection process?

The research team designed the prompts and generation procedure. The released 
`rephrased_query` paraphrases were generated by the model (GPT-4o); domain 
experts in aging and technology wrote the paraphrases used as few-shot examples 
in the prompt and later performed the face-validity review.

### Over what timeframe was the data collected?

The synthetic queries were generated in 2025. The underlying diary study 
was conducted earlier; see the paper, Section 3.

### Were any ethical review processes conducted?

Yes. All study procedures for the underlying diary study were approved by the
Institutional Review Board (IRB) of the University of Illinois Chicago. STAQ
itself is synthetic and derived from that IRB-approved study.

### Does the dataset relate to people?

Indirectly. STAQ contains no real individuals' data, but it is designed to
represent the communication of older adults and was grounded in data collected
from older-adult participants under IRB approval.

## Preprocessing, cleaning, and labeling

### Was any preprocessing, cleaning, or labeling done?

Communication-characteristic labels were assigned as part of generation. The
generated text was analyzed for token length and lexical diversity (type-token
ratio) to validate realism against the real data. For any additional filtering or
curation steps, see the paper.

### Was the raw data saved in addition to the processed data?

The real diary-study queries that seeded generation are kept private to protect
participant privacy and are not released. The released STAQ file is the
distributed artifact.

### What software or models were used?

GPT-4o was used for generation. A Sentence-BERT (SBERT) model with Principal
Component Analysis (PCA) was used to validate semantic fidelity against the real
data.

## Uses

### Has the dataset been used for any tasks already?

No; STAQ is released as a resource and was not used for a downstream task in the paper.

### What other tasks could the dataset be used for?

Potential uses include training or evaluating age-inclusive query understanding,
paraphrasing, and clarification systems, and research on accessible technology
support for older adults.

### Is there anything about the composition or collection that might affect future uses?

Yes. The data is synthetic (en-US) and grounded in one diary study; it may not
generalize across languages, cultures, or the full diversity of older adults.
`expert_rating` covers only the 50-query sample. Users should account for these
limits and validate against real data where possible.

### Are there tasks for which the dataset should not be used?

STAQ should not be treated as a substitute for real older adults' behavior, used
to make claims about real individuals or subgroups, or used to replace human
evaluation and participatory research.

## Distribution

### How will the dataset be distributed?

The canonical copy is archived on Zenodo with a Digital Object Identifier (DOI),
with a mirror on Hugging Face (`hhshomee2/STAQ`) for discoverability.

### Under what license?

The dataset is released under the Creative Commons Attribution 4.0 International
license (CC BY 4.0): free to share and adapt with attribution.

### Are there IP-based or other restrictions?

No restrictions beyond the attribution requirement of CC BY 4.0. No third-party
IP, export controls, or usage fees apply.

## Maintenance

### Who maintains the dataset, and how can they be contacted?

The dataset is maintained by the authors. See the [README](README.md) for the
current contact and citation details.

### Will the dataset be updated, and how will updates be communicated?

Updates and corrections will be released as new versions on Zenodo, each with its
own version DOI under a shared concept DOI, so earlier versions remain citable and
accessible.

---

Version [1.0.0] — [date]. Datasheet schema: Gebru, T., Morgenstern, J., Vecchione,
B., Vaughan, J. W., Wallach, H., Daumé III, H., & Crawford, K. (2021). Datasheets
for Datasets. *Communications of the ACM*, 64(12), 86–92.