High-quality human data, from your spreadsheet
Azmaval turns a CSV or JSON file into a labelling screen, then gets it reviewed by vetted people with consensus, gold checks and agreement figures built in.
How it works
Paste your data
CSV, TSV, JSON, JSON Lines or a Label Studio export. Up to 25 MB and 20,000 rows.
Check the screen
Azmaval reads the columns, proposes the review screen and shows a live preview. Edit anything it guessed wrong.
Get it reviewed
Label it yourself in private mode, or invite vetted reviewers. Consensus, gold checks and agreement are built in.
Export clean data
JSONL, CSV, Parquet, Label Studio JSON or CoNLL, or pull results through the REST API.
Your data decides the screen
There is no template to pick and no config to write. A file like this becomes a sentiment question with the text shown and the id kept out of sight:
id,text
1,The delivery was early and the box was undamaged
2,Nothing arrived and nobody answered my emails
3,It works as describedEvery kind of question
A sample can carry up to four questions on one screen.
Built for quality and for languages others skip
Quality you can inspect
Gold checks, majority consensus, agreement per sample and reviewer accuracy with statistical bounds, so small samples do not promote or punish anyone early.
Vetted reviewers
Reviewers join by invitation, start on probation and can be traced to whoever vouched for them. Or keep a campaign private and label alone.
Pashto, Dari, Urdu and Arabic
Reviewer screens in all four, laid out right to left, and guidance for the script and dialect issues that trip up annotation.
Works on any phone
Reviewers install the web app from the browser with no app store, and answers save offline and sync when the connection returns.
Moving from Label Studio
Paste an export and the questions are rebuilt. Results export back as Label Studio JSON, with a labelling config.
Comparisons
Where Label Studio, Argilla and Scale AI are stronger, and where Azmaval is, with sources.
Questions
What is Azmaval?
Azmaval is a data annotation platform. You paste a CSV or JSON file, it builds a review screen from the columns, and vetted reviewers label the data. You export clean results with agreement figures.
Do I need to write a labelling config?
No. Azmaval infers what each column is and proposes the screen. You can edit it in a form and see a live preview of the real reviewer screen.
What kinds of annotation are supported?
Single and multi-label classification, yes or no, rating scales, pairwise preference, free text, text spans with relations, bounding boxes, polygons, keypoints and audio or video segments. A sample can carry up to four questions.
Can I label my own data alone?
Yes. A private campaign is visible only to you, needs one review per sample and has no payments.
Which languages are supported?
Reviewer screens are available in English, Pashto, Persian, Urdu and Arabic, with right-to-left layouts. Data can be in any language.
Can I move from Label Studio?
Yes. Paste a Label Studio JSON export and the questions are rebuilt. Results export back as Label Studio JSON.
Is Azmaval open to everyone?
Not yet. Azmaval is in early access and invite-only, so reviewers are vetted. Ask for access from the contact page.