Abstract:
Despite recent progress, Optical Character Recognition (OCR) on historical newspapers still leaves 5–10% character errors. We present a fully automated post-OCR normalization pipeline that combines lightweight 7–8B instruction-tuned LLMs quantized to 4-bit (INT4) with a small set of regex rules. On the BLN600 benchmark (600 pages of 19th-century British newspapers), our best model YandexGPT-5-Instruct Q4 reduces Character Error Rate (CER) from 8.4% to 4.0% (–52.5%) and Word Error Rate (WER) from 20.2% to 6.5% (–67.8%), while raising semantic similarity to 0.962. The system runs on consumer hardware (RTX-4060 Ti, 8 GB VRAM) at about 35 seconds per page and requires no fine-tuning or parallel training data. These results indicate that compact INT4 LLMs are a practical alternative to large checkpoints for post-OCR cleanup of historical documents.