SOLUTIONS · FOR AI PRODUCTS
Pre-label training data with your own model
Pre-label text and pictures using your reviewed examples. Export labels and confidence scores, and focus annotation effort on uncertain cases.
Labelling 50,000 rows by hand is slow and expensive, and most of the rows are easy. If your team has already labeled a few hundred, StayCharted trains a model on them, pre-labels the rest with a confidence, and helps your annotators focus on unsure rows. No labeling platform or data scientist needed: it starts from a spreadsheet or a ZIP of pictures.
How it saves work
- Your team labels a first batch, a few hundred rows across all labels.
- StayCharted trains a model and reports how often it agrees with your annotators on rows it never saw.
- It fills the rest, each row with a label and a confidence.
- Annotators prioritize rows below your cutoff, in the Review Queue or in the spreadsheet, sorted by confidence.
- Their corrections train the next version, which you evaluate before using on the next batch.
Compare each new version on reviewed examples. Corrections can improve the next version, but improvement is not guaranteed.
The Review Queue is available on Essentials and Business for supported app and API workflows. Google Sheets fills do not automatically enter that queue; review those predictions in the sheet. Each filled row counts toward your allowance.
Haven't labeled anything yet?
Start from rows with no labels. Give StayCharted your label names, and it groups similar rows and suggests a label for each. Your team confirms or corrects, and it learns from every row it's shown. The confirmed rows become the first training set. This is on the Essentials and Business plans.
Built for clean training data
The same checks that protect StayCharted's own models help your dataset:
- Duplicates and near-duplicates, to help you find examples that could leak between training and test sets;
- The same row with two labels, which is usually an annotator disagreement worth settling;
- Labels with too few examples;
- Detected personal information, which you can mask before training;
- For pictures: unreadable files, near-identical shots, and pictures that look like they belong to another label.
What it works with
| Data | Examples |
|---|---|
| Text, one or several columns | Support messages, product titles, documents, survey answers |
| Pictures | Product photos, inspection photos, damage photos (with Vision) |
| Pictures with text | A photo with the customer's message |
| Answers with several parts | A list of labels, or two fields such as category and priority |
What it doesn't do
- It doesn't label spans, boxes or segments. It labels whole rows and whole pictures: classification, not entity or object annotation.
- It isn't a labeling platform. There's no annotator management or inter-annotator-agreement tooling. It pre-labels and shows you what to check.
- Pre-labels can bias annotators. Have reviewers check a random sample of the confident rows too, not only the unsure ones.
Evaluate before automating
Use reviewed examples that reflect your current rules. Check errors by category and sample confident answers as well as uncertain ones. Confidence is a model score, not a guarantee. Keep mandatory safety and access rules in your application. Best practices for training and review →.
Common questions
What is pre-labeling?
Using a model to suggest labels for unlabeled data, so people check and correct them instead of labeling every row from scratch.
How many rows do I need to label first?
About 20 per label to start; a few hundred in total is common. The report shows when the model is good enough to help.
Can I export the labeled data?
Yes. Fill the file and download it with every row's label and confidence, to train any model you like.
Does it work for images?
Yes, with the Vision add-on: upload pictures as a ZIP of folders, or as a spreadsheet of links.
Is our data used to train anyone else's model?
No. Models and data stay in your workspace. See Security.