SynthNet Search

The data are primarily sourced from the APA PsycTests Database with permission from the American Psychological Association (APA). The database includes approximately 30,000 scale representations derived from around 500,000 individual items that were extracted by parsing PDFs using (visual) language models.

Model accuracy: The accuracy of the magnolia-psychometrics/surveybot3000 model used for item-based search is reported in detail in the sources referenced below (Hommel & Arslan, 2025). In brief, when predicting scale correlations, the synthetic model estimates converge with empirical correlations from human respondents with an accuracy of r = .83 (95% CI [.81; .85]; manifest).

For scale-based search using construct labels, please see Wulff & Mata (2025) on the accuracy of the pre-trained model (sentence-transformers/all-mpnet-base-v2).

Data trustworthiness: While we strive for high data quality, please note that the items and scales are extracted automatically from source PDFs. As such, there may be occasional discrepancies or errors in the extracted text. We recommend verifying any critical information against the original source when possible. The extraction process involved two models: a visual language model (Qwen/Qwen2.5-VL-7B-Instruct) to parse the PDF layout and a large language model (Qwen/Qwen3-32B) to interpret and structure the content. Errors in transcription may occur, especially with complex survey layouts or poor scan quality. Warning codes are provided to flag potential issues (see below).

Warning codes indicate potential data quality issues detected during automated AI-extraction from source PDFs. While these codes do not necessarily imply errors, they highlight areas that may require additional scrutiny. Regardless of warning codes, we recommend verifying any critical information against the original source when possible.

In item-based search, cosine similarity scores are closely aligned with the correlation coefficients one would expect when administering two scales to human respondents. In scale-based search, similarity scores reflect semantic relatedness between the search term and the scale and/or instrument labels in the database.

While cosine similarity theoretically ranges from -1 to 1, sentence vectors rarely yield negative scores because their many abstract linguistic features are predominantly positively correlated.

Unfortunately, due to copyright restrictions, we cannot display individual survey items (statements/questions) directly on the platform. However, you can follow the DOI link to obtain the items from the original source. To access full instrument details, please:

• Click the DOI link in search results to access the original source
• Contact the instrument authors for permission to use the full measure
• Check institutional access through your library's APA PsycTests subscription