We explore LLM applications for page stream segmentation in insurance document processing, demonstrating that parameter-efficient fine-tuning achieves strong accuracy but revealing significant calibration challenges that limit deployment confidence.
We create TabMe++, an enhanced page stream segmentation benchmark with commercial-grade OCR, and show that parameter-efficiently fine-tuned decoder-based LLMs like Mistral-7B achieve 80% straight-through processing rates, outperforming encoder-based models.
A computational social science project that built a 47,000+ bill dataset from Congress.gov (115th-117th Congresses), with a co-sponsorship legislative graph and TF-IDF baseline models for 33-class policy-area classification (up to ~0.89 weighted F1 on full text), now available on Hugging Face.