LabelLoop Label a little. Learn a lot. Nothing leaves your device.

What is this?

Small teams collect piles of open-text feedback - survey comments, program reviews, suggestion boxes - and then read every single one by hand, or pay for a cloud service that takes the data away. LabelLoop is a third option: active learning in your browser. You label a few comments, a small model trains on your device, and it shows you the comments it is most unsure about next - often the ones worth a look first. Suggestions focus your review; they are not a guarantee.

Everything runs locally in this tab: the text, the training, the model. There is no server. Save your work as a file and resume anywhere.

Try the demo

360 synthetic comments in the style of feedback for a community education nonprofit - six themes, realistic noise, some genuinely ambiguous ones. Ground truth is hidden so you can label for real and see how you agree.

Bring your own feedback

Upload a CSV with a text column (optional id and label columns). The file is read locally in your browser and never uploaded anywhere.

Resume from a project file

No project loaded

Progress

Load a project first.

What the model has learned

Strongest words and phrases per theme, by model weight. This is how you check the model is listening to the right signals.

Unlabeled pile, predicted

Where the model would put the rest, and how sure it is. Review before trusting.

Recent activity

    Does active learning actually help?

    We simulated the labeling loop on the demo dataset (360 synthetic comments): start with 2 labels per theme, then repeatedly ask for 5 more - picked either by the model (active) or at random - retrain, and measure macro-F1 on a held-out 30% split the simulated labeler never sees. Three seeds; the band is ±1 standard deviation. The numbers below are computed by this repository's code (run_benchmark.mjs), not written by hand.

    Held-out quality with all labels

    These results come from a synthetic demo dataset with clean themes. Real feedback is messier and scores will be lower. Treat this as evidence that the labeling loop works end to end - not as a promised accuracy for your data.

    Save & resume

    Your project is a single JSON file: the comments, your labels, and the full history. Download it, resume it anywhere - it is the source of truth. (This tab also autosaves to your browser's local storage.)

    Export results

    Start over