Interactive Tutorial
Model Lab Demo
An animated walkthrough of every step you take on the Model Lab page. Controls on the left let you step through manually or let it auto-play.
Overview
Labeled Candidates is the number that drives model quality. You need at least 20 balanced labels (mix of Positive and Negative) before training produces a reliable model. Aim for 50+.
Train Variants
Training uses human-reviewed candidate rows only. Leave the job filter blank to use all labeled candidates stored, or enter comma-separated Job IDs to scope training to a specific batch.
Keyword-Aware Model
Uses text keywords (e.g. "comments", "reply", "posted by") combined with structural HTML features. Best accuracy for UGC detection.
Structure-Only Model
Uses only DOM shape signals — element depth, sibling counts, ARIA roles. No text keywords. Useful as a baseline to measure keyword contribution.
Trained Models
Use Runtime JSON when the browser extension needs a compact deployable inference bundle. Click Use to select a model for scoring jobs or the site-group probe.
| Artifact | Variant | Created | Precision | Recall | F1 | Top-1 |
|---|
Score Existing Job
Use a trained model against a completed job. This scores every stored candidate, re-ranks each page, and shows whether the top candidate is confident or needs manual review.
Site Group Probe
Paste or upload URLs or hostnames. The probe matches them against pages you have already scanned, scores with the selected model, and groups results by hostname.