Short-form Text Rewriting with Phi Silica
Researchers have explored adapting a small language model, Phi Silica, for the specific task of short-form text rewriting. They curated a dataset from presentation slides and used GPT-5 for generating rewrites and evaluations. The study found that fine-tuning Phi Silica improved its semantic accuracy, reduced hallucinations, and made it more competitive against GPT-5's rewrites. AI
IMPACT Demonstrates methods for improving small language model performance on precision-critical tasks, potentially enabling more efficient on-device AI.