Stuck at 70% Accuracy: Preventing LLMs from Altering Numbers in PDF Translation

Nervous_Classroom714 · reddit · 2026-08-31

A developer discusses a bottleneck in their PDF translation tool where the LLM occasionally reformats numbers embedded in sentences (e.g., decimal flipping, localization), despite pinning tokens. They tried extracting from source PDFs, small batches, and glossary passes. The post seeks advice on prompting patterns for in-sentence numbers, balancing batch size vs. context, and handling echoed segments.

Original post →

More from coding & agent

coding & agent channel →