Run a research program
Meet the duties of a research program on your instance, export a consented research dataset, and record removals and closure.
A research program uses the data of people who allowed research use on the first-login screen. Use this procedure to export a structured research dataset, record the removals that you owe, and close a dataset. The optional research question describes how to turn on the question and what it asks.
Before you start
- You need the instance admin role and your own session. By default, each action also needs a recent passkey confirmation, even if admin access needs no passkey.
config.researchOrganizationnames your research organization.- You meet every duty in What operators must do, before the first export and while the research runs.
A dataset is a JSON file of structured practice history for the participants whose current consent permits it. It is pseudonymous personal data, not anonymous data. Keep it within the research team. It leaves out the text of participants and models, source code, credentials, and provider identities.
Export a dataset
- Open Instance administration → Research datasets.
- If the page asks for it, select Confirm with passkey.
- Under Export a dataset, choose the Workspace.
- Enter First day (UTC) and Last day (UTC). The window includes the last day, can span up to 400 days, and ends no later than today.
- Enter Use allowed through (UTC), from the study protocol.
- Select Export and download.
- Save the file and note its dataset ID.
Hephaestus records the export and a protected link to its participants, but keeps no copy of the file. You cannot download it again. A new export creates a new dataset. The export cannot change the research purpose, the organization, or the notice version.
Record a removal
Before each new analysis, check whether you owe a removal.
- Under Open datasets, select Check open datasets.
- Read the participants under Why removal is owed, for example Withdrew consent or Account erased.
- Remove each listed participant key's rows from every copy that is not anonymized. Also remove each review fact that lists the key. If you cannot, delete the copy.
- Select Record removal, choose the participants under Participants removed, and confirm.
An empty list does not mean that no dataset exists. A later consent of the same participant does not cancel a removal that you owe.
Close a dataset
Stop using a dataset on its Use allowed through day.
- Delete or anonymize every copy of the file and of the data taken from it.
- Select Close dataset on the dataset.
- Choose Every copy was deleted or Every copy was anonymized, and confirm.
Closure removes the link to the participants and keeps the record of the export. A recorded removal or closure is your statement. Hephaestus cannot reach or revoke a downloaded copy.
Check that it worked
- Select Check open datasets.
- Check that each dataset says No removal is owed, or that a closed dataset no longer appears.
What operators must do
The broad consent is valid only if you meet these duties:
- Before each export, analysis, or benchmark run, use only participants with a current grant. The grant must name the current organization and cover the current scope. Research datasets applies this rule to each export.
- Remove the data of a participant who withdraws from datasets that are not anonymized, without undue delay. Treat an erasure request and a Slack message-use opt-out the same way.
- Keep participants informed about the research purpose, methods, projects, and released datasets. Use a research information page, or a named contact who sends updates (EDPB 05/2020, paragraphs 161 and 162).
- Assess whether the research needs an ethics committee vote, and get it where your institution requires one. Do not state an approval that you do not hold.
- Keep the research team separate from grading and line management, because a power imbalance can weaken a free choice.
- Test a dataset for re-identification before you call it anonymous. Free text and code are hard to anonymize.
- Honor the AI choice of each participant in benchmark runs. Model providers see only pseudonymized data, under a data processing agreement.
- Do not use the consent for browsing activity from the browser extension.
- Keep each safeguard that the consent text states. Replace identifiers with a code, store the key apart, and screen text for names. Give access only to the team and its processors.
Publish the retention and the recipients of the research in your privacy notice.
Datasets outside the team
The consent has no second choice for sharing or publishing datasets. Only anonymized data leaves the research team. Anonymous data is outside the GDPR (Recital 26). A release needs a new consent wording in these cases:
- It contains data that is not anonymized.
- It serves a purpose outside the area of research.
- It goes to a third country without adequate safeguards.