
Scope & Challenge: Built a DeBERTa-based classifier to distinguish human-written from AI-generated text. The key challenge was severe model overfitting during the fine-tuning phase.
Deliverables & Feedback: Architected a stabilized training pipeline with curriculum-based sampling and momentum-guided updates. Delivered a robust model achieving 92.5% accuracy, outperforming rule-based and LLM-based classifier baselines, which now serves as a highly reliable baseline for the client's anti-plagiarism workflows.
1 like · 1 view · Sep 15, 2026
Scope & Challenge: Built a DeBERTa-based classifier to distinguish human-written from AI-generated text.