Domain-Adapted Small Language Models with Hybrid Post-Processing: Achieving Cost-Efficient, Low-Latency Multi-Label Structured Prediction via LoRA Fine-Tuning on Scarce Data

Srinivasan Manoharan, Dilipkumar Nallusamy, Sachin Kumar +1

Deploying frontier large language models (LLMs) for domain-specific structured evaluation tasks incurs prohibitive latency, cost, and data-privacy overhead. We present a hybrid fra…