STAYCHARTED AMT / INSIGHTS

How AI Model Trainer works behind the scenes

Explore the design decisions behind AMT’s text classifiers and personal-information checks.

01

Choose text features using validation rows

Distinctive words can identify a part code or a category-specific phrase, while meaning can connect different descriptions of the same request. AI Classifier Models fit meaning alone and words plus meaning, then retain the candidate that scores better on held-back validation rows. On BANKING77, words plus meaning reached 92.6% accuracy, compared with 91.0% for meaning alone. On Medical Abstracts, meaning alone reached 60.0%, compared with 52.0% for the combination. There is no universal best representation: evaluate the selection on independent test examples before relying on it.

02

Compare fine-tuning with a lighter classifier

A Dedicated AI Model fine-tunes the AI itself on labeled examples. In the assistant comparison, it reached 93.7% on 3,080 BANKING77 test messages and 84.9% on 4,000 LEDGAR clauses. Claude, given category definitions without labeled examples, reached 84.7% and 76.9%. The medical results were closer: Claude 65.7% and Dedicated 63.6%. These workflows received different training inputs; the comparison does not establish that fine-tuning always wins.

03

Measure picture representations on the task

An AI Image Classifier Model uses a fixed visual representation and learns your categories from it. SigLIP 2 reached 91.1% on our product-photo benchmark and 95.2% on pet breeds; DINOv2 base did better on the breeds at 96.0%. This is why we report dataset-specific results rather than one ranking for every visual task. Dedicated image training is another approach to evaluate when subtle category differences matter.

04

Look past overall accuracy

The cashew benchmark illustrates the cost of relying on one headline number. The Dedicated image model found 26 of 40 defective items (65%); the AI Image Classifier found 16 (40%). A review cutoff helps select uncertain rows, but confident mistakes remain possible. Our projected review results assume people correct every flagged mistake without introducing errors. They are not measured results from a human-review study.

05

Mask supported patterns before learning

Masking replaces supported personal-information patterns with labels before training and applies the same transformation when that model reads new text. It does not detect every name, address or sensitive free-text detail. It also does not erase original uploads or remove information learned by an older version. Inspect the detected values, remove other unnecessary sensitive content and retrain when the input policy changes.

EXPLORE NEXT

Try it on your own examples.

Free includes Classifier Models and one AI Classifier Model. Compare plans for more AI capacity, Dedicated AI Models, Vision, and API access.