This is the full developer documentation for Peekbank
# Getting Started
> How to get started with Peekbank
#### Welcome to Peekbank!
These docs will set you up so that you can work with the data offered by peekbank.\
Pick what you want to do:
* Access the data › R: Data Access with PeekbankR
* Access the data › Other Languages: Data Access with Redivis
* Learn about the paradigm: [Looking while listening](https://pmc.ncbi.nlm.nih.gov/articles/PMC9950292/#S1)
* Understand how the data is structured: Data Schema, Codebook
* Share my own study data: Sharing Data
* Check docs for contributors: Task Index, Dataset Lifecycle
# Data Access - PeekbankR
> A guide on how to access peekbank data using peekbankr.
The `peekbankr` package offers a set of wrappers that enable you to access the peekbank data in R.
## Installation
```r
# install.packages("remotes")
remotes::install_github("peekbank/peekbankr")
```
## Authorising Redivis
You need a free [Redivis account](https://redivis.com). The first data request opens a browser window asking you to authenticate.
I have an environment that cannot authenticate via browser
Use a personal API token instead. Create one in your [Redivis workspace](https://redivis.com/workspace) settings, give it the `data.metadata` scope, and put it in your `.Renviron`:
```plaintext
REDIVIS_API_TOKEN=your-token-here
```
## Basic Usage
Create a client for a release and read the tables from it:
```r
library(peekbankr)
library(dplyr)
pb <- peekbank("2026.2")
data <- pb$aoi_timepoints |>
left_join(pb$administrations)
```
Every table in the data schema is a field on the client.\
They are downloaded on first access and cached for the session.
## Pinning a Release
`peekbank()` requires you to specify a release:
```r
peekbank_releases() # available releases
pb <- peekbank("2026.2")
pb$release # "2026.2"
```
Alternatively, `"current"` gets you the most recently published version. Avoid this in contexts where reproducibility is important, as “current” might point to a different release by the time the code is rerun.
See Releases for what each release contains.
## Subsetting
To download only part of a table, `pb$subset$()` takes `dplyr` expressions and returns only the matching rows:
```r
aoi_timepoints <- pb$subset$aoi_timepoints(dataset_name == "pomper_saffran_2016")
ps_data <- aoi_timepoints |>
left_join(pb$administrations)
```
Predicates can name columns of the table and of anything a row is linked to:
```r
pb$subset$aoi_timepoints(age >= 24, age <= 36, sex == "female")
pb$subset$aoi_timepoints(vanilla_trial, !excluded)
pb$subset$administrations(sex == "female")
pb$subset$trials(vanilla_trial)
pb$subset$aoi_timepoints(dataset_name == "pomper_saffran_2016",
age < 24, vanilla_trial)
```
For the timepoint tables specifically, the predicates are translated to remote SQL queries to limit transfer size, meaning a subset of predicates is available (the [basic filter functions](https://dplyr.tidyverse.org/reference/filter.html#useful-filter-functions) plus [Redivis’ additions](https://docs.redivis.com/api/client-libraries/redivis-r/reference/redivis/redivisusdconnect_dbi#supported-dplyr-functions)). The functions emit error messages to inform you if you try to use an untranslatable predicate.
## Server-side Queries
`pb$db` holds a [dbplyr](https://dbplyr.tidyverse.org/) handle on each table. `dplyr` verbs on it compile to SQL and run on the server; data transfer happens when you `collect()`. Here is an example that summarises a timepoint table without pulling it:
```r
library(dplyr)
pb$db$aoi_timepoints |>
count(aoi) |>
collect()
```
## PeekbankR API
The PeekbankR API page covers every table and method on the client, and the helper functions.
# Data Access - Redivis
> A guide on how to access peekbank data outside of R.
If you are not using R, you can read the tables directly from [peekbank on Redivis](https://stanford.redivis.com/datasets/a3v0-3drqa89kg), or browse and query them there in the browser.
## Versions
Every Redivis version of the dataset holds one Peekbank release, named in its `release_info` table.
Which Redivis version holds which release
| Release | Redivis version |
| --- | --- |
| 2026.2 | v2.1 |
| 2026.1 | v1.4 |
| 2025.2 | v1.3 |
| 2025.1 | v1.2 |
| 2022.1 | v1.1 |
| 2021.1 | v1.0 |
| 2021.0 | v2.0 |
## Clients
Reading a table with the Python client:
```python
import redivis # uv add redivis
table = (
redivis.organization("datapages")
.dataset("peekbank:a3v0", version="v2.1")
.table("administrations")
)
df = table.to_pandas_dataframe()
```
Authentication uses a `REDIVIS_API_TOKEN` environment variable. For queries and the JavaScript client, see the [Redivis client documentation](https://docs.redivis.com/api/client-libraries).
Differences to peekbankr
peekbankr reshapes some of the Redivis tables before handing them out, creating minor differences to what you would get with direct access:
* peekbankr fetches the `aoi_timepoints` table in the compressed run-length encoding. Accessing the raw table via the Redivis client is a larger transfer, so either run your queries on Redivis before fetching or pull `aoi_timepoints_rle` directly and expand the table yourself.
* peekbankr adds a `dataset_name` column to `administrations`, `subjects`, `trials`, `trial_types` and `stimuli`.
* `trials.excluded` and `trial_types.vanilla_trial` are 0/1 on Redivis, while peekbankr transforms them to TRUE/FALSE.
# External Links
> Where to find Peekbank's data, tools, papers and code.
## Data
* [Peekbank on Redivis](https://stanford.redivis.com/datasets/a3v0-3drqa89kg), the released data tables
* [peekbank\_files on Redivis](https://stanford.redivis.com/datasets/frvk-5c922jwkx), raw and processed data
* [Peekbank datasets sheet](https://docs.google.com/spreadsheets/d/1nGXWRu6_q7ATGrdWrFDuDmb9VoGHCnTO9NLhqhpuN28/edit?usp=sharing), a list of datasets with import status
* Interactive analyses for exploring the data in the browser
## Repositories
* [peekbankr](https://github.com/peekbank/peekbankr), source code and issue tracker for the R package
* [peekbank-data-import](https://github.com/peekbank/peekbank-data-import), the scripts that convert each raw dataset into the Peekbank format, and the issue tracker for problems in the data
* [peekbank-website](https://github.com/peekbank/peekbank-website), the code for this site
## Papers
* Boyce et al. (2026). Data-driven recommendations for increasing reliability and validity in measures of infant word recognition. *PsyArXiv*. [Preprint](https://osf.io/preprints/psyarxiv/28gq4) · [Analysis code](https://github.com/peekbank/peekbank-method)
* Frank et al. (2025). Continuous developmental changes in word recognition support language learning across early childhood. *eLife*. [Paper](https://doi.org/10.7554/eLife.109636) · [Analysis code](https://github.com/peekbank/peekbank-development)
* Zettersten et al. (2023). Peekbank: An open, large-scale repository for developmental eye-tracking data of children's word recognition. *Behavior Research Methods*. [Paper](https://doi.org/10.3758/s13428-022-01906-4) · [Analysis code](https://github.com/peekbank/peekbank-paper/tree/main/brm)
* Zettersten et al. (2021). Peekbank: Exploring children's word recognition through an open, large-scale repository for developmental eye-tracking data. *CogSci*. [Preprint](https://psyarxiv.com/ep693) · [Analysis code](https://github.com/peekbank/peekbank-paper/tree/main/cogsci2021)
# PeekbankR API
> The peekbankr client and helper functions, and how to use them.
An overview of the peekbankr client and helpers. For full argument lists, see the [full function reference](https://peekbank.github.io/peekbankr/reference/index.html).
## Creating a Client
`peekbank()` returns a client pinned to one Peekbank release, through which most of peekbankr is used.
```r
pb <- peekbank("2026.2")
```
`peekbank_releases()` lists the releases you can pass. `"current"` takes the most recently published one.
A client caches the data it downloads for the duration of the R session. `cache_on_disk = TRUE` additionally saves the data in a `.pbcache` folder in the working directory, which persists across sessions to prevent redownloading.
```r
pb <- peekbank("2026.2", cache_on_disk = TRUE)
```
## Tables
The client has one field per table in the data schema, each reachable three ways (here `trials`):
| | |
| --- | --- |
| `pb$trials` | the whole table, fetched once per session and reused after that |
| `pb$subset$trials(...)` | narrowed by `dplyr` predicates |
| `pb$db$trials` | a `dbplyr` handle, transferred on `collect()` |
In addition to the fields described in the schema, `subjects`, `administrations`, `trials`, `trial_types` and `stimuli` also carry a `dataset_name` column when returned by peekbankr.
### Whole tables
`pb$` returns the whole table as a tibble. It is downloaded once on first access and kept for the rest of the R session for subsequent use.
```r
pb$datasets
```
```r
pb$subjects
```
```r
pb$administrations
```
```r
pb$trials
```
```r
pb$trial_types
```
```r
pb$stimuli
```
```r
pb$aoi_region_sets
```
```r
pb$aoi_timepoints
```
```r
pb$xy_timepoints
```
### Subsets
`pb$subset$(...)` returns only the rows matching the `dplyr` predicates you pass. You can pass multiple predicates separated by comma, which are then combined with AND. A predicate can name the table’s own columns and the columns of anything a row is uniquely linked to:
`datasets`
Own columns:
`dataset_id`, `lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`subjects`
Own columns:
`subject_id`, `sex`, `native_language`, `lab_subject_id`, `subject_aux_data`
From `datasets` :
`dataset_id`, `lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`administrations`
Own columns:
`administration_id`, `dataset_id`, `age`, `lab_age`, `lab_age_units`, `subject_id`, `monitor_size_x`, `monitor_size_y`, `sample_rate`, `tracker`, `coding_method`, `administration_aux_data`
From `subjects` :
`sex`, `native_language`, `lab_subject_id`, `subject_aux_data`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`trials`
Own columns:
`trial_id`, `trial_order`, `trial_type_id`, `excluded`, `exclusion_reason`, `trial_aux_data`
From `trial_types` :
`aoi_region_set_id`, `dataset_id`, `distractor_id`, `target_id`, `full_phrase`, `full_phrase_language`, `point_of_disambiguation`, `target_side`, `condition`, `lab_trial_id`, `vanilla_trial`, `trial_type_aux_data`
From `stimuli` (target):
`target_stimulus_novelty`, `target_original_stimulus_label`, `target_english_stimulus_label`, `target_stimulus_image_path`, `target_lab_stimulus_id`, `target_image_description`, `target_image_description_source`, `target_stimulus_aux_data`
From `stimuli` (distractor):
`distractor_stimulus_novelty`, `distractor_original_stimulus_label`, `distractor_english_stimulus_label`, `distractor_stimulus_image_path`, `distractor_lab_stimulus_id`, `distractor_image_description`, `distractor_image_description_source`, `distractor_stimulus_aux_data`
From `aoi_region_sets` :
`l_x_max`, `l_x_min`, `l_y_max`, `l_y_min`, `r_x_max`, `r_x_min`, `r_y_max`, `r_y_min`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`trial_types`
Own columns:
`trial_type_id`, `aoi_region_set_id`, `dataset_id`, `distractor_id`, `target_id`, `full_phrase`, `full_phrase_language`, `point_of_disambiguation`, `target_side`, `condition`, `lab_trial_id`, `vanilla_trial`, `trial_type_aux_data`
From `stimuli` (target):
`target_stimulus_novelty`, `target_original_stimulus_label`, `target_english_stimulus_label`, `target_stimulus_image_path`, `target_lab_stimulus_id`, `target_image_description`, `target_image_description_source`, `target_stimulus_aux_data`
From `stimuli` (distractor):
`distractor_stimulus_novelty`, `distractor_original_stimulus_label`, `distractor_english_stimulus_label`, `distractor_stimulus_image_path`, `distractor_lab_stimulus_id`, `distractor_image_description`, `distractor_image_description_source`, `distractor_stimulus_aux_data`
From `aoi_region_sets` :
`l_x_max`, `l_x_min`, `l_y_max`, `l_y_min`, `r_x_max`, `r_x_min`, `r_y_max`, `r_y_min`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`stimuli`
Own columns:
`stimulus_id`, `dataset_id`, `stimulus_novelty`, `original_stimulus_label`, `english_stimulus_label`, `stimulus_image_path`, `lab_stimulus_id`, `image_description`, `image_description_source`, `stimulus_aux_data`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
`aoi_timepoints`
Own columns:
`administration_id`, `trial_id`, `aoi`, `t_norm`
From `administrations` :
`dataset_id`, `age`, `lab_age`, `lab_age_units`, `subject_id`, `monitor_size_x`, `monitor_size_y`, `sample_rate`, `tracker`, `coding_method`, `administration_aux_data`
From `subjects` :
`sex`, `native_language`, `lab_subject_id`, `subject_aux_data`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
From `trials` :
`trial_order`, `trial_type_id`, `excluded`, `exclusion_reason`, `trial_aux_data`
From `trial_types` :
`aoi_region_set_id`, `distractor_id`, `target_id`, `full_phrase`, `full_phrase_language`, `point_of_disambiguation`, `target_side`, `condition`, `lab_trial_id`, `vanilla_trial`, `trial_type_aux_data`
From `stimuli` (target):
`target_stimulus_novelty`, `target_original_stimulus_label`, `target_english_stimulus_label`, `target_stimulus_image_path`, `target_lab_stimulus_id`, `target_image_description`, `target_image_description_source`, `target_stimulus_aux_data`
From `stimuli` (distractor):
`distractor_stimulus_novelty`, `distractor_original_stimulus_label`, `distractor_english_stimulus_label`, `distractor_stimulus_image_path`, `distractor_lab_stimulus_id`, `distractor_image_description`, `distractor_image_description_source`, `distractor_stimulus_aux_data`
From `aoi_region_sets` :
`l_x_max`, `l_x_min`, `l_y_max`, `l_y_min`, `r_x_max`, `r_x_min`, `r_y_max`, `r_y_min`
`xy_timepoints`
Own columns:
`xy_timepoint_id`, `administration_id`, `trial_id`, `x`, `y`, `t_norm`
From `administrations` :
`dataset_id`, `age`, `lab_age`, `lab_age_units`, `subject_id`, `monitor_size_x`, `monitor_size_y`, `sample_rate`, `tracker`, `coding_method`, `administration_aux_data`
From `subjects` :
`sex`, `native_language`, `lab_subject_id`, `subject_aux_data`
From `datasets` :
`lab_dataset_id`, `dataset_name`, `shortcite`, `cite`, `dataset_aux_data`
From `trials` :
`trial_order`, `trial_type_id`, `excluded`, `exclusion_reason`, `trial_aux_data`
From `trial_types` :
`aoi_region_set_id`, `distractor_id`, `target_id`, `full_phrase`, `full_phrase_language`, `point_of_disambiguation`, `target_side`, `condition`, `lab_trial_id`, `vanilla_trial`, `trial_type_aux_data`
From `stimuli` (target):
`target_stimulus_novelty`, `target_original_stimulus_label`, `target_english_stimulus_label`, `target_stimulus_image_path`, `target_lab_stimulus_id`, `target_image_description`, `target_image_description_source`, `target_stimulus_aux_data`
From `stimuli` (distractor):
`distractor_stimulus_novelty`, `distractor_original_stimulus_label`, `distractor_english_stimulus_label`, `distractor_stimulus_image_path`, `distractor_lab_stimulus_id`, `distractor_image_description`, `distractor_image_description_source`, `distractor_stimulus_aux_data`
From `aoi_region_sets` :
`l_x_max`, `l_x_min`, `l_y_max`, `l_y_min`, `r_x_max`, `r_x_min`, `r_y_max`, `r_y_min`
`aoi_region_sets`
Own columns:
`aoi_region_set_id`, `l_x_max`, `l_x_min`, `l_y_max`, `l_y_min`, `r_x_max`, `r_x_min`, `r_y_max`, `r_y_min`
The result has only the table’s own columns and is cached for the session.
**The timepoint tables** are filtered on the server before data transfer, limiting predicates to what can be translated to SQL (the [basic filter functions](https://dplyr.tidyverse.org/reference/filter.html#useful-filter-functions) plus [Redivis’ additions](https://docs.redivis.com/api/client-libraries/redivis-r/reference/redivis/redivisusdconnect_dbi#supported-dplyr-functions)). Local variables are filled in automatically, but if you want to evaluate a function before sending it to the server (e.g. to compare some external values), you need to prefix with `!!`, e.g. `age < !!max(local_ages)`, as the function would otherwise be sent to the server for translation.
All other tables are filtered locally and take any `dplyr` expression.
### Server-side queries
`pb$db$` is a lazy `dbplyr` table on Redivis. `dplyr` verbs on it compile to SQL and run on the server, and the results are fetched on calling `collect()`.
```r
pb$db$aoi_timepoints |>
dplyr::count(aoi) |>
dplyr::collect()
```
`pb$db` also holds `aoi_timepoints_rle`, a compressed form of `aoi_timepoints`: one row per run of a constant AOI, starting at `t_norm` and `length` samples long.
## Client Fields
| Field | What it holds |
| --- | --- |
| `pb$release` | the Peekbank release this client is pinned to, e.g. `"2026.2"` |
| `pb$redivis_tag` | the Redivis version holding it, e.g. `"v2.1"` |
| `pb$tables` | the names of the tables in this release |
Printing the client shows the release it is pinned to and what is on it:
```r
pb
#> release 2026.2 (Redivis version v2.1)
#> code: https://github.com/peekbank/peekbank-data-import/tree/
#> tables: datasets subjects administrations trials trial_types stimuli
#> aoi_region_sets aoi_timepoints xy_timepoints
#> methods: clear_cache() images() query() readmes()
#> subset: $subset$()
#> db: $db$ for a dbplyr handle, read on collect()
```
## Methods
### pb$query()
Runs BigQuery Standard SQL against the pinned release and returns a tibble. String comparison is case-sensitive. Results are cached for the session.
```r
pb$query("SELECT dataset_name, shortcite FROM datasets")
```
### pb$images()
Downloads the stimulus images to `//` and returns a tibble with:
* one row per stimulus that has an image, in the requested datasets
* the columns of `pb$stimuli`, plus `local_image_path`: the path the image was saved to, or `NA` where the download failed
| Argument | Default | |
| --- | --- | --- |
| `datasets` | all datasets | names of the datasets to download images for |
| `dir` | "stimulus\_data" | directory to save them to |
| `skip_existing` | TRUE | skip images already on disk |
```r
stimuli <- pb$images(datasets = "reflook_v4")
```
### pb$readmes()
Downloads the per-dataset READMEs, which record import decisions and dataset quirks, to `/.md` and returns a tibble with:
* one row per requested dataset in the release
* the columns of `pb$datasets`, plus `local_readme_path`: the path the README was saved to, and `readme`: its text (both `NA` where the download failed)
| Argument | Default | |
| --- | --- | --- |
| `datasets` | all datasets | names of the datasets to download READMEs for |
| `dir` | "dataset\_readmes" | directory to save them to |
```r
readmes <- pb$readmes(datasets = "pomper_saffran_2016")
```
### pb$clear\_cache()
Drops this release’s cached tables and query results. A `cache_on_disk` client also deletes the saved files.
```r
pb$clear_cache()
```
## Helper Functions
### peekbank\_releases()
Returns a tibble with:
* one row per release `peekbank()` can be pinned to, most recently published first
* the columns `release`, `redivis_tag` (the Redivis version holding it) and `current` (`TRUE` for the one `"current"` resolves to)
```r
peekbank_releases()
```
### unnest\_aux()
Turns one key of a table’s `*_aux_data` JSON into rows and returns a tibble.
E.g., for `pb$subjects`, a subject with three CDI scores gets three rows in the output, each with that subject’s columns plus the fields of one score. Subjects without the key are dropped. A key holding a single value, like `native_language_non_iso`, becomes one column named after it.
See codebook for compatible versions and keys.
| Argument | Default | |
| --- | --- | --- |
| `df` | required | a table with a `*_aux_data` column, e.g. `pb$subjects` |
| `key` | required | the key to expand, e.g. "cdi\_responses" |
```r
cdi <- pb$subjects |>
unnest_aux("cdi_responses")
```
Example: Unnesting CDI JSON
This subject in `pb$subjects`
| subject\_id | sex | native\_language | lab\_subject\_id | dataset\_id | dataset\_name |
| --- | --- | --- | --- | --- | --- |
| 1836 | female | eng | 6356 | 25 | fmw\_2013 |
has this `subject_aux_data`:
```json
{
"cdi_responses": [
{"instrument_type": "ws", "measure": "prod", "age": 18,
"rawscore": 138, "percentile": 61, "language": "English (American)"},
{"instrument_type": "ws", "measure": "prod", "age": 24,
"rawscore": 313, "percentile": 45, "language": "English (American)"}
]
}
```
Running the code example above turns it into two rows, one per CDI score (with this unnesting being repeated for all subjects):
| subject\_id | sex | … | instrument\_type | measure | age | rawscore | percentile | language |
| --- | --- | - | --- | --- | --- | --- | --- | --- |
| 1836 | female | … | ws | prod | 18 | 138 | 61 | English (American) |
| 1836 | female | … | ws | prod | 24 | 313 | 45 | English (American) |
### add\_cdi\_percentiles()
Looks up where each CDI score falls in the vocabulary norms for its language, form, measure and age, and returns the scores with:
* `norm_age`: the age whose norms were used (input ages are rounded to the nearest month)
* `norm_percentile`: the highest norm percentile (5, 10, …, 95, 99) the score reaches, or `1` if below the 5th
* `norm_percentile_sex`: the same against children of the same sex (`NA` for sexes other than female and male)
The norms cover English (American) only:
* WG and WG short form, comprehension and production, 8 to 18 months
* WS and WS short form, production, 16 to 30 months
WG and WS use the 2022 American English norms, the short forms the MB-CDI short-form norms: Level I for WG short and Level IIA for WS short.
| Argument | Default | |
| --- | --- | --- |
| `cdi` | required | CDI scores from `unnest_aux(pb$subjects, "cdi_responses")` |
Example usage:
```r
cdi <- pb$subjects |>
unnest_aux("cdi_responses") |>
add_cdi_percentiles()
```
### add\_cdi\_proportions()
Divides each CDI score by the number of words on its form, taken from the form’s item list in Wordbank, and returns the scores along with:
* `n_items`: the number of words on the form
* `proportion`: the raw score divided by it
Forms without a Wordbank item list get `NA`.
| Argument | Default | |
| --- | --- | --- |
| `cdi` | required | CDI scores from `unnest_aux(pb$subjects, "cdi_responses")` |
The CDI helpers can be chained on the scores from `unnest_aux()`:
```r
cdi <- pb$subjects |>
unnest_aux("cdi_responses") |>
add_cdi_percentiles() |>
add_cdi_proportions()
```
### nearest\_cdi()
Matches each row of a table with a `subject_id` and an `age`, e.g. `pb$administrations`, to the child’s CDI scores closest to it in age from a provided cdi table, and returns:
* one row per matched score, in the input table’s order
* the columns of the input table and the columns the table supplied as the `cdi` input argument, with the score’s age as `cdi_age`
Each row gets assigned the closest score of every language, form and measure within `max_age_diff` months. Rows without any are dropped.
| Argument | Default | |
| --- | --- | --- |
| `df` | required | a table with `subject_id` and `age`, e.g. `pb$administrations` |
| `cdi` | required | CDI scores from `unnest_aux(pb$subjects, "cdi_responses")` |
| `max_age_diff` | required | the furthest a cdi administration’s age may be from the lwl administration’s age, in months |
| `prefer_later` | FALSE | use the later of two equally close scores instead of the earlier one |
```r
cdi <- pb$subjects |>
unnest_aux("cdi_responses") |>
add_cdi_percentiles()
pb$administrations |>
nearest_cdi(cdi, max_age_diff = 3)
```
Example: Matching CDI scores to administrations
These two administrations in `pb$administrations`
| administration\_id | subject\_id | age | … | dataset\_name |
| --- | --- | --- | - | --- |
| 1488 | 936 | 18 | … | fernald\_marchman\_2012 |
| 1490 | 936 | 31 | … | fernald\_marchman\_2012 |
belong to a child with these scores in `cdi`:
| subject\_id | … | language | instrument\_type | measure | age | rawscore |
| --- | - | --- | --- | --- | --- | --- |
| 936 | … | English (American) | wg | comp | 18 | 83 |
| 936 | … | English (American) | wg | prod | 18 | 17 |
| 936 | … | English (American) | ws | prod | 22 | 36 |
| 936 | … | English (American) | ws | prod | 24 | 85 |
Running the code example above turns them into two rows. The administration at 18 months gets the WG comprehension and production scores from 18 months, but no WS score, as the closest one is 4 months away. The administration at 31 months has no score within 3 months, so it is dropped:
| administration\_id | subject\_id | age | … | language | instrument\_type | measure | cdi\_age | rawscore | … |
| --- | --- | --- | - | --- | --- | --- | --- | --- | - |
| 1488 | 936 | 18 | … | English (American) | wg | comp | 18 | 83 | … |
| 1488 | 936 | 18 | … | English (American) | wg | prod | 18 | 17 | … |
# Codebook
> Every column in the Peekbank schema, table by table.
Every column in the Peekbank schema. See Data Schema for how the tables fit together. The schema changed with release 2025.1; pick the tab for the release you are using. A **?** after a type means the value can be NA.
* 2025.1 onward
## datasets
| Field | Type | Description |
| --- | --- | --- |
| `dataset_id` key | integer | row identifier for the datasets table indexing from zero |
| `lab_dataset_id` | string? | lab-internal short name of the dataset |
| `dataset_name` | string | the name used for that dataset in Peekbank |
| `shortcite` | string | shortened citation (author (year) APA-style in-text citation) |
| `cite` | string | full citation (APA format) |
| `dataset_aux_data` | json? | optional additional information about the dataset in the format of a JSON (see Auxiliary data below for the keys in use) |
## subjects
| Field | Type | Description |
| --- | --- | --- |
| `subject_id` key | integer | row identifier for the subjects table indexing from zero |
| `sex` | string | sex of the participants. One of `female`, `male`, `other`, `unspecified`. |
| `native_language` | string | participants' native language as an ISO 639-2 code; bilingual participants can have several codes, separated by ", " |
| `lab_subject_id` | string | lab unique participant identifier |
| `subject_aux_data` | json? | optional additional information about the subject in the format of a JSON (see Auxiliary data below for the keys in use) |
## administrations
| Field | Type | Description |
| --- | --- | --- |
| `administration_id` key | integer | row identifier for the administrations table indexing from zero |
| `dataset_id` → datasets | integer | ID of the dataset |
| `age` | float? | age of the participant in months, converted from the age the lab recorded: days: days / (365.25/12); months: unchanged; decimal years: years \* 12; whole years: years \* 12 + 6. |
| `lab_age` | float? | age of the participant as recorded by the lab, in lab\_age\_units |
| `lab_age_units` | string | unit of lab\_age. One of `days`, `months`, `years`. |
| `subject_id` → subjects | integer | ID of the subject |
| `monitor_size_x` | integer? | width of the monitor (in pixels) |
| `monitor_size_y` | integer? | height of the monitor (in pixels) |
| `sample_rate` | float? | sampling rate of the eyetracker (in Hz) / frame rate for manual gaze coding |
| `tracker` | string? | name of the eye-tracker type/ data source (SMI/ iCoder/ Tobii/ etc.) |
| `coding_method` | string | method used in the experiment for coding gaze data. One of `eyetracking` (gaze recorded by an eye tracker), `manual gaze coding` (looking coded by hand, frame by frame, from video), `automated gaze coding` (looking coded from video by software), `preprocessed eyetracking` (eyetracking with only AOI data available). |
| `administration_aux_data` | json? | optional additional information about the administration in the format of a JSON (see Auxiliary data below for the keys in use) |
## trials
| Field | Type | Description |
| --- | --- | --- |
| `trial_id` key | integer | row identifier for the trials table indexing from zero |
| `trial_order` | integer | index of the trial in order of presentation for this participant during the experiment |
| `trial_type_id` → trial\_types | integer | ID of the trial type |
| `excluded` | boolean | whether or not the trial was excluded in the data analysis conducted in the original experiment for which the data was collected |
| `exclusion_reason` | string? | the reason the trial was excluded in the original experiment that is the source of the data |
| `trial_aux_data` | json? | optional additional information about the trial in the format of a JSON (see Auxiliary data below for the keys in use) |
## trial\_types
| Field | Type | Description |
| --- | --- | --- |
| `trial_type_id` key | integer | row identifier for the trial\_types table indexing from zero |
| `aoi_region_set_id` → aoi\_region\_sets | integer? | ID of the AOI region set |
| `dataset_id` → datasets | integer | ID of the dataset |
| `distractor_id` → stimuli | integer | ID of the distractor stimulus |
| `target_id` → stimuli | integer | ID of the target stimulus |
| `full_phrase` | string? | Full phrase prompting the target |
| `full_phrase_language` | string | language of the full phrase: an ISO 639-2/B code, or multiple, artificial or other. The codes are listed in the 639-2/B column of |
| `point_of_disambiguation` | integer | target onset time (relative to trial onset) in ms |
| `target_side` | string | part of the screen in which the target stimulus appears. One of `left`, `right`. |
| `condition` | string? | condition of the trial, worded as closely as possible to the original study |
| `lab_trial_id` | string? | trial label assigned within lab (or descriptive label of the trial) |
| `vanilla_trial` | boolean | whether or not the trial is a "standard" ("vanilla") looking-while-listening trial or includes additional manipulations. Some requirements for being considered a standard/ vanilla trial: Familiar words (also no part-words); Both target and distractor are familiar objects (i.e., each could be valid familiar targets; non-prototypical objects are OK, but unnatural/ artificially modified exemplars are not); Target word is the first point of disambiguation; Well-formed/grammatical carrier phrase; No relevant “stuff”/information prenominally (e.g. semantically informative verbs, adjectives); No nonsense words anywhere (including carrier phrase); No language/ speaker/ accent/ etc. switching within trial; No duplicated target words; No intentional background noise or audio filtering. |
| `trial_type_aux_data` | json? | optional additional information about the trial type in the format of a JSON (see Auxiliary data below for the keys in use) |
## stimuli
| Field | Type | Description |
| --- | --- | --- |
| `stimulus_id` key | integer | row identifier for the stimuli table indexing from zero |
| `dataset_id` → datasets | integer | ID of the dataset |
| `stimulus_novelty` | string | whether the stimulus word (not image) is novel or familiar. One of `novel`, `familiar`. |
| `original_stimulus_label` | string? | label of the stimulus in the original language of the study: the word used for the image when it is the target, or its label in the experiment design |
| `english_stimulus_label` | string? | label assigned to the stimulus translated into English |
| `stimulus_image_path` | string? | path to the image file, relative to raw\_data/; NA if the raw data has no image file |
| `lab_stimulus_id` | string? | an id or name assigned to the stimulus by the original authors |
| `image_description` | string? | a short natural language description of the image |
| `image_description_source` | string | source of the natural language description. One of `image path` (taken from the image path), `experiment documentation` (based on experiment documentation), `Peekbank discretion` (generated by the importer/ Peekbank team). |
| `stimulus_aux_data` | json? | optional additional information about the stimulus in the format of a JSON (see Auxiliary data below for the keys in use) |
## aoi\_timepoints
| Field | Type | Description |
| --- | --- | --- |
| `aoi_timepoint_id` key | integer | row identifier for the aoi\_timepoints table indexing from zero |
| `aoi` | string | name of the looking location. One of `target` (looking at the target), `distractor` (looking at the distractor), `other` (on-screen looking, but not in the AOI regions for target or distractor), `missing` (missing data, e.g. an off-screen look, blink or tracker error). |
| `administration_id` → administrations | integer | ID of the administration |
| `t_norm` | integer | timestamp in ms, centered on the point of disambiguation of the target (= 0 ms); samples are 25 ms apart (40 Hz) |
| `trial_id` → trials | integer | ID of the trial |
## xy\_timepoints
| Field | Type | Description |
| --- | --- | --- |
| `xy_timepoint_id` key | integer | row identifier for the xy\_timepoints table indexing from zero |
| `administration_id` → administrations | integer | ID of the administration |
| `trial_id` → trials | integer | ID of the trial |
| `x` | integer? | x-coordinate of gaze position at timepoint t (origin (0,0) is bottom left) |
| `y` | integer? | y-coordinate of gaze position at timepoint t (origin (0,0) is bottom left) |
| `t_norm` | integer | timestamp in ms, centered on the point of disambiguation of the target (= 0 ms); samples are 25 ms apart (40 Hz) |
## aoi\_region\_sets
| Field | Type | Description |
| --- | --- | --- |
| `aoi_region_set_id` key | integer | row identifier for the aoi\_region\_sets table indexing from zero |
| `l_x_max` | integer | maximum x-coordinate of the left AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `l_x_min` | integer | minimum x-coordinate of the left AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `l_y_max` | integer | maximum y-coordinate of the left AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `l_y_min` | integer | minimum y-coordinate of the left AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `r_x_max` | integer | maximum x-coordinate of the right AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `r_x_min` | integer | minimum x-coordinate of the right AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `r_y_max` | integer | maximum y-coordinate of the right AOI region (with the bottom left corner of the screen as origin (0,0)) |
| `r_y_min` | integer | minimum y-coordinate of the right AOI region (with the bottom left corner of the screen as origin (0,0)) |
## Auxiliary data
The `*_aux_data` columns hold JSON with the keys below. The top-level keys are optional; the indented fields sit inside nested JSON objects.
### subject\_aux\_data
| Field | Type | Description |
| --- | --- | --- |
| `cdi_responses` | JSON | JSON object for CDI responses |
| `cdi_responses.instrument_type` | string | CDI version/instrument used (wg, ws, wgshort or wsshort) |
| `cdi_responses.measure` | string | CDI measure (comprehension or production) |
| `cdi_responses.rawscore` | numeric | CDI raw vocabulary score |
| `cdi_responses.percentile` | numeric | CDI percentile (as reported in the raw data) |
| `cdi_responses.age` | numeric | age of the participant in months when the instrument was administered |
| `cdi_responses.language` | string | CDI language (in Wordbank format) |
| `lang_exposures` | JSON | JSON object for language exposures |
| `lang_exposures.language` | string | language that a child is exposed to (in Wordbank format) |
| `lang_exposures.exposure` | numeric | percentage exposure in a given language |
| `lang_measures` | JSON | Scores on language measures other than the CDI |
| `lang_measures.instrument_type` | string | The specific language development measure used (e.g., LUI, LDS, DVAP) |
| `lang_measures.rawscore` | numeric | raw score in the language measure instrument |
| `lang_measures.age` | numeric | age of the participant in months when the instrument was administered |
| `lang_measures.language` | string | language of the measure (in Wordbank format) |
| `native_language_non_iso` | string | name of participants' native language if the language is not represented in ISO 639-2 |
### trial\_type\_aux\_data
| Field | Type | Description |
| --- | --- | --- |
| `full_phrase_language_non_iso` | string | language of the full phrase if the language is not represented in ISO 639-2 |
* 2021.0 – 2022.1
These releases have the same tables and fields as 2025.1 onward, except that these fields do not exist: `datasets.dataset_aux_data`, `subjects.subject_aux_data`, `administrations.administration_aux_data`, `trials.excluded`, `trials.exclusion_reason`, `trials.trial_aux_data`, `trial_types.vanilla_trial`, `trial_types.trial_type_aux_data`, `stimuli.stimulus_aux_data`.
# Data Schema
> The tables in the Peekbank schema and the IDs that link them.
The Peekbank schema is a set of tables linked by IDs, each accessible through peekbankr. For details on the fields of each table, refer to the codebook.
If you need a refresher on tables, keys and joins, the [Joins chapter of R for Data Science](https://r4ds.hadley.nz/joins.html) covers them.
The schema changed with release 2025.1; pick the tab for the release you are using.
* 2025.1 onward
| Table | One row is |
| --- | --- |
| `datasets` | a contributed dataset, with its citation |
| `subjects` | a participant |
| `administrations` | one session a participant took part in, with their age and the tracker used |
| `trials` | a trial a participant completed, in order, with whether the original study excluded it |
| `trial_types` | a type of trial, defined by its target and distractor stimuli, target side and carrier phrase |
| `stimuli` | a word–image pair |
| `aoi_timepoints` | a 25 ms gaze sample: target, distractor, other or missing |
| `xy_timepoints` | a 25 ms gaze sample as screen coordinates (where available) |
| `aoi_region_sets` | the screen positions of the left and right areas of interest (where available) |
* 2021.0 – 2022.1
| Table | One row is |
| --- | --- |
| `datasets` | a contributed dataset, with its citation |
| `subjects` | a participant |
| `administrations` | one session a participant took part in, with their age and the tracker used |
| `trials` | a trial type at one position in a dataset, shared by every participant who saw it there |
| `trial_types` | a type of trial, defined by its target and distractor stimuli, target side and carrier phrase |
| `stimuli` | a word–image pair |
| `aoi_timepoints` | a 25 ms gaze sample: target, distractor, other or missing |
| `xy_timepoints` | a 25 ms gaze sample as screen coordinates (where available) |
| `aoi_region_sets` | the screen positions of the left and right areas of interest (where available) |
# Releases
> The Peekbank database releases.
We release new versions of the Peekbank database as new data becomes available. To ensure reproducibility, we keep past major database versions available. Pass the version name to `peekbank()` as a string to pin it for your analysis. Editors of the dataset on Redivis can read the unreleased draft of the next release with `peekbank("next")`.
#### 2026.2
Latest full release: 44 datasets, 3,917 children.
#### 2026.1
Peekbank version for reproducing the preprint of Boyce et al. (2026), *PsyArXiv*, .
#### 2025.1
Peekbank version for reproducing the preprint version of Frank et al. (2025), *eLife*, .
#### 2022.1
Peekbank version for reproducing Zettersten et al. (2023), *Behavior Research Methods*, .
#### 2021.0
Peekbank version for reproducing Zettersten et al. (2021), *CogSci*, .
# Dataset Lifecycle
> How a dataset moves through the Peekbank stack, from a contributing lab to a release.
This page gives you a tour through the Peekbank stack by describing the stages a newly added dataset goes through. You can click on an element in the diagram below to jump to the related section.
* Release workflow
* Cached workflow for quick testing
## Raw data in peekbank\_files
A lab sends their LWL data (or a link to a data repository) to someone on the Peekbank team, who uploads it to [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx/files) on Redivis. Each dataset has its own folder there, named after its identifier (for example `weaver_zettersten_2024`), holding the raw files.
## Import
The import process turns the raw data into preprocessed data in the Peekbank format. For each dataset, a collaborator writes a separate, scaffolded `import.R` script in [peekbank-data-import](https://github.com/peekbank/peekbank-data-import), which downloads the raw data from [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx/files), reshapes it (outputting one `.csv` per table of the schema) and validates the result. A dataset specific `README.md` documents any import decisions.
The import scripts use some internal helper functions from [PeekbankR](https://github.com/peekbank/peekbankr).
Once the data passes validation, we perform an internal review step where another collaborator looks through the import script and the paper the data came from in tandem. Until then, a marker file in the dataset’s folder keeps it out of the pipeline and the releases. If no further issues are found, the reviewer removes the marker from the folder, which makes the dataset part of the next release.
With a flag in the import script, the processed data and `README.md` can also be uploaded to [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx/files), into a subfolder of `_processed_data` folder. Releases do not use these uploads, but the cached workflow for testing does, to allow for faster end-to-end testing of import changes.
## Ingestion
The ingestion builds a draft of the next release of [Peekbank on Redivis](https://stanford.redivis.com/datasets/a3v0-3drqa89kg) from the raw data and the import code. It is triggered via a command in [peekbank-data-import](https://github.com/peekbank/peekbank-data-import) (`./pbi ingest `), which runs in Docker and pins both the raw data and the state of the import repository to the release. It downloads the raw data, reruns every import script and then combines the processed `.csv` files of all datasets into one set of tables, offsetting the row IDs to their position in the joined dataset. The process stops if any dataset fails to import or validate. Peekbank uses release names in the format \.\.
Every dataset’s `README.md` and raw data are bundled with the release in the `dataset_files` table. These files enable us to reproduce the ingestion if needed and allow PeekbankR to fetch versioned stimulus images and READMEs. Raw data that may not be shared publicly is not uploaded here, but the release notes what version of the private data was used.
Other than the schemas main tables, the ingest also creates an `aoi_timepoints_rle` table, a [run-length-encoded](https://en.wikipedia.org/wiki/Run-length_encoding) version of `aoi_timepoints`, which is used by PeekbankR to cut transmission size. Similarly, the ingestion creates a `release_info` table with metadata, which has a single row with the release name of the dataset, the commit of [peekbank-data-import](https://github.com/peekbank/peekbank-data-import) it was built from, and the version of the private raw data it used.
Testing with the cached workflow
Rerunning every import takes a while, so for quickly testing import changes there is a quicker way, depicted in the “Cached workflow for quick testing” tab above. `./pbi ingest --cache` builds a draft from the cached `.csv` files uploaded to `_processed_data`. Such a draft records no import repository commit and is labelled as a test build on Redivis. Individuals with edit access on the Peekbank dataset on Redivis can read it with `peekbank("next")` in PeekbankR for testing purposes.
## Release
Releasing the draft with `./pbi release ` (or the Redivis interface) publishes it as a new release of the Peekbank dataset. Redivis numbers its versions on its own (e.g., `v2.1`), so our `release_info` table is used to keep track of Peekbank release names (e.g., `2026.2`). Previous versions stay available to keep old analyses reproducible. The list of supported releases can be seen on the Releases page.
## Access
End users have three ways to access the data of a release:
* PeekbankR, the R package most people use. It allows you to specify a release and read the tables from Redivis into your R environment.
* Redivis itself, where you can browse and query the tables in the browser or read them with one of its clients.
* This website, as it features interactive visualisations that let you explore some analyses on the data directly in the browser.
Furthermore, certain elements of this website (e.g., the dataset list on the Citation page) are populated with data from the releases. These, as well as the analyses, don’t query Redivis live. When the website is deployed, a Python script pulls the releases and turns them into compressed per-dataset JSON files that the browser loads.
## Corrections
When someone finds a problem in the data, it goes into the issue tracker of [peekbank-data-import](https://github.com/peekbank/peekbank-data-import). That dataset’s import script is then fixed, and the corrected data reaches everyone with the next release.
# Task Index
> For each task on Peekbank, which repository to work in and what access you need.
This list is meant to give you an overview of the parts of Peekbank you can contribute to, and what prerequisites (accounts, permissions on resources, etc.) are needed.
## Analyse Peekbank data
You need
* a free [Redivis account](https://redivis.com)
The released data is public, so just use our provided interfaces to pull it and refer to the Data Schema and Codebook for the details on the data structure. If you opt for PeekbankR, we provide helper functions to facilitate analyses involving CDI data or stimulus images.
## Share your own study data
If you have collected looking-while-listening data and want it in Peekbank, great! We will take care of most of the import work; refer to the page on Sharing Data for details.
## Add raw data to be imported
You need
* a Redivis account
* edit access to [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx) on Redivis
* edit access to the [Peekbank datasets sheet](https://docs.google.com/spreadsheets/d/1nGXWRu6_q7ATGrdWrFDuDmb9VoGHCnTO9NLhqhpuN28/edit?usp=sharing) on Google Sheets
* (optional) for datasets with private raw data, edit access to [peekbank\_private\_raw](https://stanford.redivis.com/datasets/2vqs-2cfz9magg) on Redivis
To add a dataset we have received to Peekbank’s raw data:
1. Pick an identifying internal name for it (see the list of current datasets for examples).
2. Add the dataset to the [Peekbank datasets sheet](https://docs.google.com/spreadsheets/d/1nGXWRu6_q7ATGrdWrFDuDmb9VoGHCnTO9NLhqhpuN28/edit?usp=sharing).
3. Open the draft version of [peekbank\_files](https://stanford.redivis.com/datapages/editor/datasets/frvk-5c922jwkx/files) on Redivis.
4. Click `+ Upload files` (clicking `Create` if asked to create a new version), choose `Upload directory` and select the local folder containing the raw data.
5. In the `Destination folder` dropdown, scroll down and click `+ Create new folder`.
6. Enter the internal name from step 1 under `New folder name` and click `Upload`. Files in nested folders keep their parent folders as a prefix.
Making newly added raw data available to importers outside the Peekbank team
The added files now live in the draft `"next"` version. By default, only people with `edit` privileges on `datapages.peekbank_files` can read this version; everyone else gets the latest published version. If you want people without this privilege to write import scripts (e.g. in a Peekbank hackathon), use the `Review and release` button to create a versioned release.
If you want to add or replace files, use the `+ Upload files` button and choose the folder of the dataset you want to edit. If you want to delete files, click on the dataset’s folder name in the right panel and delete the file.
What about datasets that do not allow us to share the raw data?
Raw data that may not be shared publicly (e.g. the NIH Baby Toolbox dataset) goes into [peekbank\_private\_raw](https://stanford.redivis.com/datapages/editor/datasets/2vqs-2cfz9magg), which only the Peekbank team can see. To upload such a dataset in a way that the pipeline can pick up:
1. Pick an identifying internal name for it (see the list of current datasets for examples).
2. Add the dataset to the [Peekbank datasets sheet](https://docs.google.com/spreadsheets/d/1nGXWRu6_q7ATGrdWrFDuDmb9VoGHCnTO9NLhqhpuN28/edit?usp=sharing).
3. Open the draft version of [peekbank\_private\_raw](https://stanford.redivis.com/datapages/editor/datasets/2vqs-2cfz9magg) (this is the private one) on Redivis.
4. Click `+ Upload files` (clicking `Create` if asked to create a new version), choose `Upload directory` and select the local folder containing the raw data.
5. In the `Destination folder` dropdown, scroll down and click `+ Create new folder`.
6. Enter the internal name from step 1 under `New folder name` and click `Upload`. Files in nested folders keep their parent folders as a prefix.
7. On your computer, create an empty file named `raw_data_in_private_repo.txt`.
8. Open the draft version of [peekbank\_files](https://stanford.redivis.com/datapages/editor/datasets/frvk-5c922jwkx/files) (this is the public one) on Redivis.
9. Click `+ Upload files`, choose `Upload files` and select the file from step 7.
10. In the `Destination folder` dropdown, scroll down and click `+ Create new folder`.
11. Enter the same internal name from step 1 under `New folder name` and click `Upload`. This placeholder tells the import scripts to look for the raw data in peekbank\_private\_raw.
The import scripts then automatically download the raw data from `peekbank_private_raw` for everyone who has access to it. The processed data still goes to the public `peekbank_files`, as that data has the same shape as what will be distributed via the Peekbank releases.
## Transform data into the Peekbank schema
You need
* a Redivis account
* a GitHub account
* either write access to [peekbank-data-import](https://github.com/peekbank/peekbank-data-import) (ask the Peekbank team) or your own fork to open pull requests from
* (optional) to upload the result, edit access to [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx)
* (optional) for datasets with private raw data, read access to [peekbank\_private\_raw](https://stanford.redivis.com/datasets/2vqs-2cfz9magg)
To turn a raw dataset into Peekbank-ready tables (“data import”):
1. Clone [peekbank-data-import](https://github.com/peekbank/peekbank-data-import).
2. Pick a dataset from the [Peekbank datasets sheet](https://docs.google.com/spreadsheets/d/1nGXWRu6_q7ATGrdWrFDuDmb9VoGHCnTO9NLhqhpuN28/edit?usp=sharing) that hasn’t been imported yet.
3. Set up an import script for it as described in the repository’s [README](https://github.com/peekbank/peekbank-data-import#peekbank-data-import).
4. Fill in the code following the steps in the documentation, and note any relevant import decisions in the dataset’s README.
5. Run the script. It validates the data and prints a summary with plots to check against the original paper.
6. Once the data looks right, set `upload=TRUE` in `write_and_validate_list()` and run it again to upload the processed data to Redivis.
7. Ask someone on the team to review the import. Once they are happy with it, they delete `ignored_by_pipeline_remove_after_review.txt` from the dataset folder and commit the deletion along with any fixes, which adds the dataset to the next release.
## Release a new Peekbank version
You need
* a Redivis account
* edit access to [Peekbank on Redivis](https://stanford.redivis.com/datasets/a3v0-3drqa89kg)
* edit access to [peekbank\_files](https://stanford.redivis.com/datasets/frvk-5c922jwkx) on Redivis, as the release uploads the processed data that changed
* read access to [peekbank\_private\_raw](https://stanford.redivis.com/datasets/2vqs-2cfz9magg) on Redivis, as a release reruns every import (including those with private raw data)
To build and publish a release:
1. Clone [peekbank-data-import](https://github.com/peekbank/peekbank-data-import) and set it up as described in its [ingestion README](https://github.com/peekbank/peekbank-data-import/tree/master/ingest#you-need).
2. Commit hashes of that repository get baked into releases. If you have any local changes, commit them, get them onto master on GitHub and/or pull the latest master. Only the folders of datasets that still hold an `ignored_by_pipeline_remove_after_review.txt` may differ from the commit.
3. Pick a release name in the format \.\, e.g. `2026.2`.
4. Run `./pbi ingest `. This reruns every import from scratch, builds the release tables and stages them in the draft of [Peekbank on Redivis](https://stanford.redivis.com/datapages/editor/datasets/a3v0-3drqa89kg). Nothing is staged if any dataset fails to import or validate.
5. Look through the draft on Redivis, or read it in PeekbankR with `peekbank("next")`. If you need to apply fixes, running the ingestion command again replaces the current draft on Redivis.
6. Release the draft with `./pbi release `. The `Review and release` button on Redivis works too, but is discouraged as it does not perform the metadata validation of the command.
7. If desired, add the release to the curated Releases page of this website.
## Extend PeekbankR
You need
* a GitHub account
* either write access to [peekbankr](https://github.com/peekbank/peekbankr) (ask the Peekbank team) or your own fork to open pull requests from
The code and the issue tracker can be found on GitHub: [peekbankr](https://github.com/peekbank/peekbankr).
## Extend this website
You need
* a GitHub account
* either write access to [peekbank-website](https://github.com/peekbank/peekbank-website) (ask the Peekbank team) or your own fork to open pull requests from
The code and the issue tracker can be found on GitHub: [peekbank-website](https://github.com/peekbank/peekbank-website).
# About
> About Peekbank
## About Peekbank
Peekbank collects looking-while-listening studies. In these, a child sees two pictures, hears one of them named, and an eye tracker or a video coder records where they look. Various labs have run this task since the late 1990s. Peekbank aims to generate one central repository for the data they have collected to answer questions that wouldn’t be feasible with single samples. In the process, we bring the heterogenous datasets into a common format and offer the data to the scientific community through a hosted interface.
Peekbank is open source and under active development. If you encounter a problem, please file an issue in the GitHub Issues page of the appropriate repository.
If you would like to contact us directly, email us at peekbank-dev\[at]lists.stanford.edu.
## Publications
* Boyce, V., Marchman, V. A., Baumgartner, H. A., Bergey, C. A., Braginsky, M., Giovannetti, F., Kachergis, G., Mankewitz, J., Meylan, S., Prystawski, B., Sparks, R. Z., Steffan, A., Tan, A. W. M., Frank, M. C., & Zettersten, M. (2026). **Data-driven recommendations for increasing reliability and validity in measures of infant word recognition.** *PsyArXiv*. [\[preprint\]](https://osf.io/preprints/psyarxiv/28gq4)
* Frank, M. C., Marchman, V. A., Bergey, C. A., Boyce, V., Braginsky, M., Kachergis, G., Mankewitz, J., Meylan, S., Prystawski, B., Ram, N., Sparks, R. Z., Steffan, A., Tan, A. W. M., & Zettersten, M. (2025). **Continuous developmental changes in word recognition support language learning across early childhood.** *eLife*.
* Zettersten, M., Yurovsky, D., Xu, T. L., Uner, S., Tsui, A.S.T., Schneider, R. M., Saleh, A. N., Meylan, S. C., Marchman, V., Mankewitz, J., MacDonald, K., Long, B., Lewis, M., Kachergis, G., Handa, K., deMayo, B., Carstensen, A., Braginsky, M., Boyce, V., Bhatt, N., Bergey, C. A., & Frank, M.C. (2023). **Peekbank: An open, large-scale repository for developmental eye-tracking data of children's word recognition.** *Behavior Research Methods, 55*, 2485-2500. [\[preprint\]](https://psyarxiv.com/tgnzv)
* Zettersten, M., Bergey, C. A., Bhatt, N., Boyce, V., Braginsky, M., Carstensen, A., deMayo, B., Kachergis, G., Lewis, M., Long, B., MacDonald, K., Mankewitz, J., Meylan, S. C., Saleh, A. N., Schneider, R. M., Tsui, A., Uner, S., Xu, T. L., Yurovsky, D., & Frank, M.C. (2021). **Peekbank: Exploring children's word recognition through an open, large-scale repository for developmental eye-tracking data.** *Proceedings of the 43rd Annual Conference of the Cognitive Science Society*. [\[preprint\]](https://psyarxiv.com/ep693)
## Meet the Peekbank team
[Claire Bergey](https://github.com/cbergey)
Naiti Bhatt
[Veronica Boyce](https://vboyce.github.io/)
[Mika Braginsky](https://mikabr.io/)
[Michael C. Frank](https://web.stanford.edu/~mcfrank/)
[Federico Giovannetti](https://fedegiovannetti.github.io/en/)
Kunal Handa
[George Kachergis](http://www.kachergis.com/)
[Molly Lewis](http://www.andrew.cmu.edu/user/mollylew/)
[Bria Long](https://www.brialong.com/)
[Kyle MacDonald](https://kemacdonald.com)
[Jessica Mankewitz](https://github.com/JMankewitz)
[Virginia Marchman](https://profiles.stanford.edu/virginia-marchman)
[Stephan Meylan](https://stephanmeylan.com/)
[Ben Prystawski](https://benprystawski.com/)
Annissa Saleh
[Rose M. Schneider](https://rosemschneider.github.io/)
[Adrian Steffan](https://adriansteffan.com)
[Alvin Tan](https://alvinwmtan.github.io/)
[Angeline Tsui](https://angelinetsui.github.io/)
Sarp Uner
[Tian Linger Xu](http://homes.sice.indiana.edu/txu/)
[Martin Zettersten](https://mzettersten.github.io/)
# How to Cite
> What to cite when you use Peekbank data, and the list of contributing datasets.
Thank you to all of the generous researchers who have contributed data to Peekbank!
1. Researchers publishing or presenting work using data in the Peekbank database must cite the publications linked to that dataset, as shown below in the contributors list. (If there is no linked citation, then no citation for the individual dataset is necessary).
2. For tracking use of the database and its affiliated tools, please also cite:
Zettersten, M., Yurovsky, D., Xu, T. L., Uner, S., Tsui, A.S.T., Schneider, R. M., Saleh, A. N., Meylan, S. C., Marchman, V., Mankewitz, J., MacDonald, K., Long, B., Lewis, M., Kachergis, G., Handa, K., deMayo, B., Carstensen, A., Braginsky, M., Boyce, V., Bhatt, N., Bergey, C. A., & Frank, M.C. (2023). Peekbank: An open, large-scale repository for developmental eye-tracking data of children’s word recognition. *Behavior Research Methods, 55*, 2485-2500.
3. If researchers use more than 5 datasets, we ask users to cite all data. But in case of severe space limitations, the database as a whole may be cited.
4. Since the Peekbank database is dynamic, we recommend that you list the date of download for your data in your manuscript. For example: “In these analyses, all available datasets in English from Peekbank were used (Zettersten et al., 2023). Data were downloaded on 1/9/2024.”
## Current Datasets
Below is a list of all datasets currently available in Peekbank. Their citations are also in the `cite` column of the `datasets` table.
44 datasets, as of release 2026.2.
# Sharing Data
> What to share with Peekbank, in what format, and what happens to it.
Thank you so much for offering to share the data from your study with Peekbank!
## What data and materials should I share?
The primary data we are interested in is
* time course data
* with gazes to a target or distractor image
* in response to a target label
* recorded
* at a regular sampling rate (e.g., AOI hits or coordinates captured via eye-tracker)
* or on a frame-by-frame basis (e.g., looking side coded by a human).
We are also interested in information about the particular trial:
* what target and distractor objects/images are on the screen
* what target word and carrier phrase infants heard
* the timing of the target word onset
* the order of the trials
Example shape of handcoded raw data
| subject | trial | target | distractor | tar\_loc | onset\_ms | time\_ms | gaze |
| --- | --- | --- | --- | --- | --- | --- | --- |
| s01 | 1 | ball | shoe | left | 2000 | 0 | right |
| s01 | 1 | ball | shoe | left | 2000 | 33 | right |
| … | … | … | … | … | … | … | … |
| s01 | 1 | ball | shoe | left | 2000 | 2367 | away |
| s01 | 1 | ball | shoe | left | 2000 | 2400 | left |
| … | … | … | … | … | … | … | … |
| s01 | 2 | dog | cup | right | 1850 | 0 | left |
| … | … | … | … | … | … | … | … |
For participants, we are currently tracking some basic data:
* age
* sex
* native language
We’d especially welcome vocabulary information, e.g. MCDI or other language measures. We also welcome you to share the stimuli (audio, images, videos) used in the experiment.
Please also share the citation/reference you would like us to include for the dataset (e.g., doi or citation in APA format).
For a full overview of the data we are tracking, see our codebook. To see examples of shared datasets, visit our [Redivis file archive](https://stanford.redivis.com/datasets/frvk-5c922jwkx).
## What format should the data be in?
Whatever format is easiest/most convenient for you to share! If multiple formats are possible, our preference is for “raw” data relative to pre-processed data.
## What will happen with my data?
Once added to the Peekbank database, your data will be publicly available on Redivis and in the Peekbank public database. The data will be available both in your original format and in the standardized Peekbank data format. Users will be able to access your data via the peekbankr API and the interactive analyses on this site.
## How do I get credit?
Our policy is that users must cite the original datasets when publishing or presenting work using the Peekbank database. Contributors to the database will also be referenced on the How to Cite page of our website.
## Get in touch
To contribute your data, [get in touch](mailto:mzettersten@ucsd.edu?subject=Contributing%20data%20to%20Peekbank).