msulemans/open-model-training-lab

Open Model Training Lab

An interactive, beginner-friendly account of training Qwen3, BERT-Large and DeBERTa-v3-large for BANKING77 intent classification on an Apple M2 Max.

The guide covers fifteen controlled experiments, including numerical failures, rejected refinements, data-leakage protections, a searchable glossary, quiz and interview practice.

Best recorded result: 92.99% validation accuracy and 94.12% reporting-only test accuracy with a DeBERTa upper-layer refinement.

Source and reproduction instructions: msulemans/open-model-training-lab

This is educational software, not a production banking classifier.