OneStop dataset

OneStop Eye Movements is a 360-participant English eye-tracking-while-reading corpus (Berzak, Malmaud, Shubi, Meiri, Lion, Levy, Scientific Data 2025, doi:10.1038/s41597-025-06272-2). The app's bundled demo is a small subset of it (2 participants, 12 paragraphs each); this page covers loading the full public corpus from OSF as a public dataset.

The corpus is 360 L1-English participants reading 30 Guardian articles (162 paragraphs, each in an Advanced and an Elementary version).

A text is one paragraph at one difficulty level

The public reports carry no unique paragraph id, so the loader composes one — {article_batch}_{article_id}_{paragraph_id}_{difficulty_level}, the same id the bundled demo ships, with the participant and the reading folded under it for the trial id. Advanced and Elementary are therefore two texts, not two renderings of one, and the app counts 330 of them across the four regimes.

That 330 is the 162 paragraphs at two levels (324), plus six more from the practice article — article_id 0, two paragraphs, Advanced only, repeated in all three batches. The published 30 articles / 162 paragraphs counts experimental material and leaves it out, so both figures are right.

Loading it

OneStop is exposed as four public datasets, one per reading regime. In the app, open Data Management and click the one you want in the list of datasets:

Dataset	What it is
OneStop — Ordinary reading	Reading for comprehension, without seeing the question first.
OneStop — Information seeking	Reading after seeing the question to be answered.
OneStop — Repeated reading	Reading a paragraph for the second time.
OneStop — Information seeking (repeated)	A second reading, after seeing the question.

Each holds every screen of its regime's trials, the reading passage and the screens around it, from the public OSF release:

Part	CLI id	What it is
Title	Title	The article title screen.
Question preview	Question_Preview	The question shown before reading (information-seeking regimes only).
Paragraph	Paragraph	The reading passage.
Question	Questions	The question shown after reading.
Answers	Answers	The four answer choices.
Feedback	Feedback	The correctness feedback screen.

The release also has a QA part (QA): the question and answers screens taken together as one interest period, not a screen of its own. Its fixations are the ones those two screens already hold, so the datasets leave it out; load it with parts=["QA"] when you want that combined period.

Every part ships an interest-area report (one row per word, with bounding boxes and reading measures) and a fixation report, all in the same schema — so each part renders as a scanpath. A trial is one reading of one paragraph, and its parts are that trial's screens, in the order they were shown — step through them with the Screen picker above the plot. Each screen keeps its own word boxes; not every reading has every part (the title screen opens an article, so only a reading of its first paragraph has one). On OSF only Paragraph is split by regime; the other parts come from one all-regimes release, which the four datasets share on disk, and each dataset keeps only its own regime's trials from them (by the reports' question_preview and repeated_reading_trial columns).

The Edit dataset screen's data-location part lists the Expected files and shows whether they're already present (until they are, the app shows the bundled demo, with a ⬇ Download now panel). If they're present the dataset loads with no network access; if not, click ⬇ Download to fetch them into the folder (cached on disk, so only the first load pays the download — reports range from tens to a few hundred MB each, and a regime has up to twelve). While it downloads, a card shows how much has arrived; Stop download ends it and deletes the partial file.

Each report is downloaded at a fixed OSF version (version 1, of 2025-05-28) and checked against that version's size, so everyone who downloads OneStop through the app gets the same files. That is also why the dataset list can show a regime's counts before you open it.

The folder defaults to the Data Management page's Download folder, where every public corpus gets its own subfolder (<folder>/OneStop). Leave that box blank for the default: data/ in a source checkout, otherwise the per-user data folder (~/.local/share/scanpath-studio/data, or %LOCALAPPDATA%\scanpath-studio\data on Windows). scanpath-studio run --download-dir DIR or SCANPATH_STUDIO_DOWNLOAD_DIR changes the default.

On a server other machines can reach

When the app is served to other machines (the hosted demo, or --server.address 0.0.0.0), the Data directory box, the folder picker and ⬇ Download are turned off: the corpus is read from the server's configured data location, and whoever runs it places the files there — or, on a trusted network, starts it with SCANPATH_LOCAL_FS=1. See Launch.

Fixation and interest-area coordinates are full-screen pixels on OneStop's 2560×1440 presentation monitor, so the canvas renders true-to-scale to that monitor.

From the Python API

The same loader is available headlessly — load_onestop returns normalized, plot-ready frames:

import scanpath_studio as sps

# Fetch + normalize the chosen regime + parts (cached under root).
words, fixations = sps.load_onestop(
    "data/OneStop",
    regime="ordinary",
    parts=["Paragraph"],  # any subset of the parts above, or QA
    download=True,
)
# The app loads every part of a regime but QA: pass
# parts=scanpath_studio.datasets.onestop_regime_parts("ordinary") for that.
pid, tid = sps.list_trials(words, fixations).iloc[0]  # or any row you want
fig = sps.plot_scanpath(words, fixations, pid, tid, canvas_size=(2560, 1440))

From the command line

scanpath-studio render --onestop data/OneStop \
    --onestop-regime ordinary --onestop-part Paragraph \
    -p <participant> -t <trial> -o out.html

--onestop-part is repeatable. Parts other than Paragraph are cut to --onestop-regime here too, as in the app.